feat: 算力引擎 Go 后端核心（new-api）

- 统一 OpenAI 兼容 /v1 中继 + 渠道/额度/令牌/流水 - 各领域包：relay 模型网关、model 数据层、controller 管理 API
2026-08-23 22:38:46 +08:00
commit 28e53d90fc
836 changed files with 165403 additions and 0 deletions
@@ -0,0 +1,703 @@
+package openai
+
+import (
+	"bytes"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"mime/multipart"
+	"net/http"
+	"net/textproto"
+	"net/url"
+	"path/filepath"
+	"strings"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	"github.com/QuantumNous/new-api/logger"
+	"github.com/QuantumNous/new-api/relay/channel"
+	"github.com/QuantumNous/new-api/relay/channel/ai360"
+	"github.com/QuantumNous/new-api/relay/channel/lingyiwanwu"
+	"github.com/QuantumNous/new-api/relaykit/dto"
+
+	//"github.com/QuantumNous/new-api/relay/channel/minimax"
+	"github.com/QuantumNous/new-api/relay/channel/openrouter"
+	"github.com/QuantumNous/new-api/relay/channel/xinference"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/QuantumNous/new-api/relay/common_handler"
+	relayconstant "github.com/QuantumNous/new-api/relay/constant"
+	"github.com/QuantumNous/new-api/relaykit/types"
+	"github.com/QuantumNous/new-api/service"
+	"github.com/QuantumNous/new-api/setting/model_setting"
+	"github.com/QuantumNous/new-api/setting/reasoning"
+	"github.com/samber/lo"
+
+	"github.com/gin-gonic/gin"
+)
+
+type Adaptor struct {
+	ChannelType    int
+	ResponseFormat string
+}
+
+func (a *Adaptor) ConvertGeminiRequest(c *gin.Context, info *relaycommon.RelayInfo, request *dto.GeminiChatRequest) (any, error) {
+	result, err := service.ConvertRequest(c, info, types.RelayFormatOpenAI, request)
+	if err != nil {
+		return nil, err
+	}
+	openaiRequest, ok := result.Value.(*dto.GeneralOpenAIRequest)
+	if !ok {
+		return nil, fmt.Errorf("expected OpenAI chat completions request, got %T", result.Value)
+	}
+	return a.ConvertOpenAIRequest(c, info, openaiRequest)
+}
+
+func (a *Adaptor) ConvertClaudeRequest(c *gin.Context, info *relaycommon.RelayInfo, request *dto.ClaudeRequest) (any, error) {
+	//if !strings.Contains(request.Model, "claude") {
+	//	return nil, fmt.Errorf("you are using openai channel type with path /v1/messages, only claude model supported convert, but got %s", request.Model)
+	//}
+	//if common.DebugEnabled {
+	//	bodyBytes := []byte(common.GetJsonString(request))
+	//	err := os.WriteFile(fmt.Sprintf("claude_request_%s.txt", c.GetString(common.RequestIdKey)), bodyBytes, 0644)
+	//	if err != nil {
+	//		println(fmt.Sprintf("failed to save request body to file: %v", err))
+	//	}
+	//}
+	result, err := service.ConvertRequest(c, info, types.RelayFormatOpenAI, request)
+	if err != nil {
+		return nil, err
+	}
+	aiRequest, ok := result.Value.(*dto.GeneralOpenAIRequest)
+	if !ok {
+		return nil, fmt.Errorf("expected OpenAI chat completions request, got %T", result.Value)
+	}
+	//if common.DebugEnabled {
+	//	println(fmt.Sprintf("convert claude to openai request result: %s", common.GetJsonString(aiRequest)))
+	//	// Save request body to file for debugging
+	//	bodyBytes := []byte(common.GetJsonString(aiRequest))
+	//	err = os.WriteFile(fmt.Sprintf("claude_to_openai_request_%s.txt", c.GetString(common.RequestIdKey)), bodyBytes, 0644)
+	//	if err != nil {
+	//		println(fmt.Sprintf("failed to save request body to file: %v", err))
+	//	}
+	//}
+	if info.SupportStreamOptions && info.IsStream {
+		aiRequest.StreamOptions = &dto.StreamOptions{
+			IncludeUsage: true,
+		}
+	}
+	return a.ConvertOpenAIRequest(c, info, aiRequest)
+}
+
+func (a *Adaptor) Init(info *relaycommon.RelayInfo) {
+	a.ChannelType = info.ChannelType
+
+	// initialize ThinkingContentInfo when thinking_to_content is enabled
+	if info.ChannelSetting.ThinkingToContent {
+		info.ThinkingContentInfo = relaycommon.ThinkingContentInfo{
+			IsFirstThinkingContent:  true,
+			SendLastThinkingContent: false,
+			HasSentThinkingContent:  false,
+		}
+	}
+}
+
+func (a *Adaptor) GetRequestURL(info *relaycommon.RelayInfo) (string, error) {
+	if info.RelayMode == relayconstant.RelayModeRealtime {
+		if strings.HasPrefix(info.ChannelBaseUrl, "https://") {
+			baseUrl := strings.TrimPrefix(info.ChannelBaseUrl, "https://")
+			baseUrl = "wss://" + baseUrl
+			info.ChannelBaseUrl = baseUrl
+		} else if strings.HasPrefix(info.ChannelBaseUrl, "http://") {
+			baseUrl := strings.TrimPrefix(info.ChannelBaseUrl, "http://")
+			baseUrl = "ws://" + baseUrl
+			info.ChannelBaseUrl = baseUrl
+		}
+	}
+	switch info.ChannelType {
+	case constant.ChannelTypeAzure:
+		apiVersion := info.ApiVersion
+		if apiVersion == "" {
+			apiVersion = constant.AzureDefaultAPIVersion
+		}
+		// https://learn.microsoft.com/en-us/azure/cognitive-services/openai/chatgpt-quickstart?pivots=rest-api&tabs=command-line#rest-api
+		requestURL := strings.Split(info.RequestURLPath, "?")[0]
+		requestURL = fmt.Sprintf("%s?api-version=%s", requestURL, apiVersion)
+		task := strings.TrimPrefix(requestURL, "/v1/")
+
+		if info.RelayFormat == types.RelayFormatClaude {
+			task = strings.TrimPrefix(task, "messages")
+			task = "chat/completions" + task
+		}
+
+		// 特殊处理 responses API（包含 compact）
+		if info.RelayMode == relayconstant.RelayModeResponses || info.RelayMode == relayconstant.RelayModeResponsesCompact {
+			responsesApiVersion := "preview"
+
+			subUrl := "/openai/v1/responses"
+			if strings.Contains(info.ChannelBaseUrl, "cognitiveservices.azure.com") {
+				subUrl = "/openai/responses"
+				responsesApiVersion = apiVersion
+			}
+
+			if info.ChannelOtherSettings.AzureResponsesVersion != "" {
+				responsesApiVersion = info.ChannelOtherSettings.AzureResponsesVersion
+			}
+
+			// compact 模式追加 /compact
+			if info.RelayMode == relayconstant.RelayModeResponsesCompact {
+				subUrl = subUrl + "/compact"
+			}
+
+			requestURL = fmt.Sprintf("%s?api-version=%s", subUrl, responsesApiVersion)
+			return relaycommon.GetFullRequestURL(info.ChannelBaseUrl, requestURL, info.ChannelType), nil
+		}
+
+		model_ := info.UpstreamModelName
+		// 2025年5月10日后创建的渠道不移除.
+		if info.ChannelCreateTime < constant.AzureNoRemoveDotTime {
+			model_ = strings.Replace(model_, ".", "", -1)
+		}
+		// https://github.com/songquanpeng/one-api/issues/67
+		requestURL = fmt.Sprintf("/openai/deployments/%s/%s", model_, task)
+		if info.RelayMode == relayconstant.RelayModeRealtime {
+			requestURL = fmt.Sprintf("/openai/realtime?deployment=%s&api-version=%s", model_, apiVersion)
+		}
+		return relaycommon.GetFullRequestURL(info.ChannelBaseUrl, requestURL, info.ChannelType), nil
+	//case constant.ChannelTypeMiniMax:
+	//	return minimax.GetRequestURL(info)
+	case constant.ChannelTypeCustom:
+		url := info.ChannelBaseUrl
+		url = strings.Replace(url, "{model}", info.UpstreamModelName, -1)
+		return url, nil
+	default:
+		if (info.RelayFormat == types.RelayFormatClaude || info.RelayFormat == types.RelayFormatGemini) &&
+			info.RelayMode != relayconstant.RelayModeResponses &&
+			info.RelayMode != relayconstant.RelayModeResponsesCompact {
+			return fmt.Sprintf("%s/v1/chat/completions", info.ChannelBaseUrl), nil
+		}
+		return relaycommon.GetFullRequestURL(info.ChannelBaseUrl, info.RequestURLPath, info.ChannelType), nil
+	}
+}
+
+func (a *Adaptor) SetupRequestHeader(c *gin.Context, header *http.Header, info *relaycommon.RelayInfo) error {
+	channel.SetupApiRequestHeader(info, c, header)
+	if info.ChannelType == constant.ChannelTypeAzure {
+		header.Set("api-key", info.ApiKey)
+		return nil
+	}
+	if info.ChannelType == constant.ChannelTypeOpenAI && "" != info.Organization {
+		header.Set("OpenAI-Organization", info.Organization)
+	}
+	// 检查 Header Override 是否已设置 Authorization，如果已设置则跳过默认设置
+	// 这样可以避免在 Header Override 应用时被覆盖（虽然 Header Override 会在之后应用，但这里作为额外保护）
+	hasAuthOverride := false
+	if len(info.HeadersOverride) > 0 {
+		for k := range info.HeadersOverride {
+			if strings.EqualFold(k, "Authorization") {
+				hasAuthOverride = true
+				break
+			}
+		}
+	}
+	if info.RelayMode == relayconstant.RelayModeRealtime {
+		// OpenAI 已下线 Realtime Beta API,GA 模型收到 beta 标识会以 beta_api_shape_disabled 拒绝;
+		// 仅对遗留 preview 模型保留 beta 标识
+		legacyRealtimeBeta := strings.Contains(info.UpstreamModelName, "-realtime-preview")
+		swp := c.Request.Header.Get("Sec-WebSocket-Protocol")
+		if swp != "" {
+			items := []string{
+				"realtime",
+				"openai-insecure-api-key." + info.ApiKey,
+			}
+			if legacyRealtimeBeta {
+				items = append(items, "openai-beta.realtime-v1")
+			}
+			header.Set("Sec-WebSocket-Protocol", strings.Join(items, ","))
+			//req.Header.Set("Sec-WebSocket-Key", c.Request.Header.Get("Sec-WebSocket-Key"))
+			//req.Header.Set("Sec-Websocket-Extensions", c.Request.Header.Get("Sec-Websocket-Extensions"))
+			//req.Header.Set("Sec-Websocket-Version", c.Request.Header.Get("Sec-Websocket-Version"))
+		} else {
+			if legacyRealtimeBeta {
+				header.Set("openai-beta", "realtime=v1")
+			}
+			if !hasAuthOverride {
+				header.Set("Authorization", "Bearer "+info.ApiKey)
+			}
+		}
+	} else {
+		if !hasAuthOverride {
+			header.Set("Authorization", "Bearer "+info.ApiKey)
+		}
+	}
+	if info.ChannelType == constant.ChannelTypeOpenRouter {
+		if header.Get("HTTP-Referer") == "" {
+			header.Set("HTTP-Referer", "https://www.newapi.ai")
+		}
+		if header.Get("X-OpenRouter-Title") == "" {
+			header.Set("X-OpenRouter-Title", "New API")
+		}
+	}
+	return nil
+}
+
+func (a *Adaptor) ConvertOpenAIRequest(c *gin.Context, info *relaycommon.RelayInfo, request *dto.GeneralOpenAIRequest) (any, error) {
+	if request == nil {
+		return nil, errors.New("request is nil")
+	}
+	if info.ChannelType != constant.ChannelTypeOpenAI && info.ChannelType != constant.ChannelTypeAzure {
+		request.StreamOptions = nil
+	}
+	if info.ChannelType == constant.ChannelTypeOpenRouter {
+		if len(request.Usage) == 0 {
+			request.Usage = json.RawMessage(`{"include":true}`)
+		}
+		// 适配 OpenRouter 的 thinking 后缀
+		if !model_setting.ShouldPreserveThinkingSuffix(info.OriginModelName) &&
+			strings.HasSuffix(info.UpstreamModelName, "-thinking") {
+			info.UpstreamModelName = strings.TrimSuffix(info.UpstreamModelName, "-thinking")
+			request.Model = info.UpstreamModelName
+			if len(request.Reasoning) == 0 {
+				reasoning := map[string]any{
+					"enabled": true,
+				}
+				if request.ReasoningEffort != "" && request.ReasoningEffort != "none" {
+					reasoning["effort"] = request.ReasoningEffort
+				}
+				marshal, err := common.Marshal(reasoning)
+				if err != nil {
+					return nil, fmt.Errorf("error marshalling reasoning: %w", err)
+				}
+				request.Reasoning = marshal
+			}
+			// 清空多余的ReasoningEffort
+			request.ReasoningEffort = ""
+		} else {
+			if len(request.Reasoning) == 0 {
+				// 适配 OpenAI 的 ReasoningEffort 格式
+				if request.ReasoningEffort != "" {
+					reasoning := map[string]any{
+						"enabled": true,
+					}
+					if request.ReasoningEffort != "none" {
+						reasoning["effort"] = request.ReasoningEffort
+						marshal, err := common.Marshal(reasoning)
+						if err != nil {
+							return nil, fmt.Errorf("error marshalling reasoning: %w", err)
+						}
+						request.Reasoning = marshal
+					}
+				}
+			}
+			request.ReasoningEffort = ""
+		}
+
+		// https://docs.anthropic.com/en/api/openai-sdk#extended-thinking-support
+		// 没有做排除3.5Haiku等，要出问题再加吧，最佳兼容性（不是
+		if request.THINKING != nil && strings.HasPrefix(info.UpstreamModelName, "anthropic") {
+			var thinking dto.Thinking // Claude标准Thinking格式
+			if err := json.Unmarshal(request.THINKING, &thinking); err != nil {
+				return nil, fmt.Errorf("error Unmarshal thinking: %w", err)
+			}
+
+			// 只有当 thinking.Type 是 "enabled" 时才处理
+			if thinking.Type == "enabled" {
+				// 检查 BudgetTokens 是否为 nil
+				if thinking.BudgetTokens == nil {
+					return nil, fmt.Errorf("BudgetTokens is nil when thinking is enabled")
+				}
+
+				reasoning := openrouter.RequestReasoning{
+					Enabled:   true,
+					MaxTokens: *thinking.BudgetTokens,
+				}
+
+				marshal, err := common.Marshal(reasoning)
+				if err != nil {
+					return nil, fmt.Errorf("error marshalling reasoning: %w", err)
+				}
+
+				request.Reasoning = marshal
+			}
+
+			// 清空 THINKING
+			request.THINKING = nil
+		}
+
+	}
+	isOModel := dto.IsOpenAIReasoningOModel(info.UpstreamModelName)
+	isGPT5Model := dto.IsOpenAIGPT5Model(info.UpstreamModelName)
+	if isOModel || isGPT5Model {
+		if lo.FromPtrOr(request.MaxCompletionTokens, uint(0)) == 0 && lo.FromPtrOr(request.MaxTokens, uint(0)) != 0 {
+			request.MaxCompletionTokens = request.MaxTokens
+			request.MaxTokens = nil
+		}
+
+		if isOModel {
+			request.Temperature = nil
+		}
+
+		// gpt-5系列模型适配 归零不再支持的参数
+		if isGPT5Model {
+			request.Temperature = nil
+			request.TopP = nil
+			request.LogProbs = nil
+		}
+
+		// 转换模型推理力度后缀
+		effort, originModel := reasoning.ParseOpenAIReasoningEffortFromModelSuffix(info.UpstreamModelName)
+		if effort != "" {
+			request.ReasoningEffort = effort
+			info.UpstreamModelName = originModel
+			request.Model = originModel
+		}
+
+		info.SetReasoningEffort(request.ReasoningEffort)
+
+		// o系列模型developer适配（o1-mini除外）
+		if !strings.HasPrefix(info.UpstreamModelName, "o1-mini") && !strings.HasPrefix(info.UpstreamModelName, "o1-preview") {
+			//修改第一个Message的内容，将system改为developer
+			if len(request.Messages) > 0 && request.Messages[0].Role == "system" {
+				request.Messages[0].Role = "developer"
+			}
+		}
+	}
+
+	return request, nil
+}
+
+func (a *Adaptor) ConvertRerankRequest(c *gin.Context, relayMode int, request dto.RerankRequest) (any, error) {
+	return request, nil
+}
+
+func (a *Adaptor) ConvertEmbeddingRequest(c *gin.Context, info *relaycommon.RelayInfo, request dto.EmbeddingRequest) (any, error) {
+	return request, nil
+}
+
+func (a *Adaptor) ConvertAudioRequest(c *gin.Context, info *relaycommon.RelayInfo, request dto.AudioRequest) (io.Reader, error) {
+	a.ResponseFormat = request.ResponseFormat
+	if info.RelayMode == relayconstant.RelayModeAudioSpeech {
+		jsonData, err := common.Marshal(request)
+		if err != nil {
+			return nil, fmt.Errorf("error marshalling object: %w", err)
+		}
+		return bytes.NewReader(jsonData), nil
+	} else {
+		var requestBody bytes.Buffer
+		writer := multipart.NewWriter(&requestBody)
+
+		writer.WriteField("model", request.Model)
+
+		formData, err2 := common.ParseMultipartFormReusable(c)
+		if err2 != nil {
+			return nil, fmt.Errorf("error parsing multipart form: %w", err2)
+		}
+
+		// 打印类似 curl 命令格式的信息
+		logger.LogDebug(c.Request.Context(), "--form 'model=\"%s\"'", request.Model)
+
+		// 遍历表单字段并打印输出
+		for key, values := range formData.Value {
+			if key == "model" {
+				continue
+			}
+			for _, value := range values {
+				writer.WriteField(key, value)
+				logger.LogDebug(c.Request.Context(), "--form '%s=\"%s\"'", key, value)
+			}
+		}
+
+		// 从 formData 中获取文件
+		fileHeaders := formData.File["file"]
+		if len(fileHeaders) == 0 {
+			return nil, errors.New("file is required")
+		}
+
+		// 使用 formData 中的第一个文件
+		fileHeader := fileHeaders[0]
+		logger.LogDebug(c.Request.Context(), "--form 'file=@\"%s\"' (size: %d bytes, content-type: %s)",
+			fileHeader.Filename, fileHeader.Size, fileHeader.Header.Get("Content-Type"))
+
+		file, err := fileHeader.Open()
+		if err != nil {
+			return nil, fmt.Errorf("error opening audio file: %v", err)
+		}
+		defer file.Close()
+
+		part, err := writer.CreateFormFile("file", fileHeader.Filename)
+		if err != nil {
+			return nil, errors.New("create form file failed")
+		}
+		if _, err := io.Copy(part, file); err != nil {
+			return nil, errors.New("copy file failed")
+		}
+
+		// 关闭 multipart 编写器以设置分界线
+		writer.Close()
+		c.Request.Header.Set("Content-Type", writer.FormDataContentType())
+		logger.LogDebug(c.Request.Context(), "--header 'Content-Type: %s'", writer.FormDataContentType())
+		return &requestBody, nil
+	}
+}
+
+func (a *Adaptor) ConvertImageRequest(c *gin.Context, info *relaycommon.RelayInfo, request dto.ImageRequest) (any, error) {
+	switch info.RelayMode {
+	case relayconstant.RelayModeImagesEdits:
+		if isJSONRequest(c) {
+			return request, nil
+		}
+
+		var requestBody bytes.Buffer
+		writer := multipart.NewWriter(&requestBody)
+
+		writer.WriteField("model", request.Model)
+		// 使用已解析的 multipart 表单，避免重复解析
+		mf := c.Request.MultipartForm
+		if mf == nil {
+			form, err := common.ParseMultipartFormReusable(c)
+			if err != nil {
+				return nil, fmt.Errorf("failed to parse multipart form: %w", err)
+			}
+			c.Request.MultipartForm = form
+			c.Request.PostForm = url.Values(form.Value)
+			mf = form
+		}
+
+		// 写入所有非文件字段
+		if mf != nil {
+			for key, values := range mf.Value {
+				if key == "model" {
+					continue
+				}
+				for _, value := range values {
+					writer.WriteField(key, value)
+				}
+			}
+		}
+
+		if mf != nil && mf.File != nil {
+			// Check if "image" field exists in any form, including array notation
+			var imageFiles []*multipart.FileHeader
+			var exists bool
+
+			// First check for standard "image" field
+			if imageFiles, exists = mf.File["image"]; !exists || len(imageFiles) == 0 {
+				// If not found, check for "image[]" field
+				if imageFiles, exists = mf.File["image[]"]; !exists || len(imageFiles) == 0 {
+					// If still not found, iterate through all fields to find any that start with "image["
+					foundArrayImages := false
+					for fieldName, files := range mf.File {
+						if strings.HasPrefix(fieldName, "image[") && len(files) > 0 {
+							foundArrayImages = true
+							imageFiles = append(imageFiles, files...)
+						}
+					}
+
+					// If no image fields found at all
+					if !foundArrayImages && (len(imageFiles) == 0) {
+						return nil, errors.New("image is required")
+					}
+				}
+			}
+
+			// Process all image files
+			for i, fileHeader := range imageFiles {
+				file, err := fileHeader.Open()
+				if err != nil {
+					return nil, fmt.Errorf("failed to open image file %d: %w", i, err)
+				}
+
+				// If multiple images, use image[] as the field name
+				fieldName := "image"
+				if len(imageFiles) > 1 {
+					fieldName = "image[]"
+				}
+
+				// Determine MIME type based on file extension
+				mimeType := detectImageMimeType(fileHeader.Filename)
+
+				// Create a form file with the appropriate content type
+				h := make(textproto.MIMEHeader)
+				h.Set("Content-Disposition", fmt.Sprintf(`form-data; name="%s"; filename="%s"`, fieldName, fileHeader.Filename))
+				h.Set("Content-Type", mimeType)
+
+				part, err := writer.CreatePart(h)
+				if err != nil {
+					return nil, fmt.Errorf("create form part failed for image %d: %w", i, err)
+				}
+
+				if _, err := io.Copy(part, file); err != nil {
+					return nil, fmt.Errorf("copy file failed for image %d: %w", i, err)
+				}
+
+				// 复制完立即关闭，避免在循环内使用 defer 占用资源
+				_ = file.Close()
+			}
+
+			// Handle mask file if present
+			if maskFiles, exists := mf.File["mask"]; exists && len(maskFiles) > 0 {
+				maskFile, err := maskFiles[0].Open()
+				if err != nil {
+					return nil, errors.New("failed to open mask file")
+				}
+				// 复制完立即关闭，避免在循环内使用 defer 占用资源
+
+				// Determine MIME type for mask file
+				mimeType := detectImageMimeType(maskFiles[0].Filename)
+
+				// Create a form file with the appropriate content type
+				h := make(textproto.MIMEHeader)
+				h.Set("Content-Disposition", fmt.Sprintf(`form-data; name="mask"; filename="%s"`, maskFiles[0].Filename))
+				h.Set("Content-Type", mimeType)
+
+				maskPart, err := writer.CreatePart(h)
+				if err != nil {
+					return nil, errors.New("create form file failed for mask")
+				}
+
+				if _, err := io.Copy(maskPart, maskFile); err != nil {
+					return nil, errors.New("copy mask file failed")
+				}
+				_ = maskFile.Close()
+			}
+		} else {
+			return nil, errors.New("no multipart form data found")
+		}
+
+		// 关闭 multipart 编写器以设置分界线
+		writer.Close()
+		c.Request.Header.Set("Content-Type", writer.FormDataContentType())
+		return &requestBody, nil
+
+	default:
+		return request, nil
+	}
+}
+
+func isJSONRequest(c *gin.Context) bool {
+	if c == nil || c.Request == nil {
+		return false
+	}
+	return strings.HasPrefix(c.Request.Header.Get("Content-Type"), "application/json")
+}
+
+// detectImageMimeType determines the MIME type based on the file extension
+func detectImageMimeType(filename string) string {
+	ext := strings.ToLower(filepath.Ext(filename))
+	switch ext {
+	case ".jpg", ".jpeg":
+		return "image/jpeg"
+	case ".png":
+		return "image/png"
+	case ".webp":
+		return "image/webp"
+	default:
+		// Try to detect from extension if possible
+		if strings.HasPrefix(ext, ".jp") {
+			return "image/jpeg"
+		}
+		// Default to png as a fallback
+		return "image/png"
+	}
+}
+
+func (a *Adaptor) ConvertOpenAIResponsesRequest(c *gin.Context, info *relaycommon.RelayInfo, request dto.OpenAIResponsesRequest) (any, error) {
+	//  转换模型推理力度后缀
+	effort, originModel := reasoning.ParseOpenAIReasoningEffortFromModelSuffix(request.Model)
+	if effort != "" {
+		if request.Reasoning == nil {
+			request.Reasoning = &dto.Reasoning{
+				Effort: effort,
+			}
+		} else {
+			request.Reasoning.Effort = effort
+		}
+		request.Model = originModel
+	}
+	if info != nil && request.Reasoning != nil && request.Reasoning.Effort != "" {
+		info.SetReasoningEffort(request.Reasoning.Effort)
+	}
+	return request, nil
+}
+
+func (a *Adaptor) DoRequest(c *gin.Context, info *relaycommon.RelayInfo, requestBody io.Reader) (any, error) {
+	if info.RelayMode == relayconstant.RelayModeAudioTranscription ||
+		info.RelayMode == relayconstant.RelayModeAudioTranslation ||
+		(info.RelayMode == relayconstant.RelayModeImagesEdits && !isJSONRequest(c)) {
+		return channel.DoFormRequest(a, c, info, requestBody)
+	} else if info.RelayMode == relayconstant.RelayModeRealtime {
+		return channel.DoWssRequest(a, c, info, requestBody)
+	} else {
+		return channel.DoApiRequest(a, c, info, requestBody)
+	}
+}
+
+func (a *Adaptor) DoResponse(c *gin.Context, resp *http.Response, info *relaycommon.RelayInfo) (usage any, err *types.NewAPIError) {
+	switch info.RelayMode {
+	case relayconstant.RelayModeRealtime:
+		err, usage = OpenaiRealtimeHandler(c, info)
+	case relayconstant.RelayModeAudioSpeech:
+		usage = OpenaiTTSHandler(c, resp, info)
+	case relayconstant.RelayModeAudioTranslation:
+		fallthrough
+	case relayconstant.RelayModeAudioTranscription:
+		err, usage = OpenaiSTTHandler(c, resp, info, a.ResponseFormat)
+	case relayconstant.RelayModeImagesGenerations, relayconstant.RelayModeImagesEdits:
+		if info.IsStream {
+			usage, err = OpenaiImageStreamHandler(c, info, resp)
+		} else {
+			usage, err = OpenaiImageHandler(c, info, resp)
+		}
+	case relayconstant.RelayModeRerank:
+		usage, err = common_handler.RerankHandler(c, info, resp)
+	case relayconstant.RelayModeResponses:
+		if info.IsStream {
+			usage, err = OaiResponsesStreamHandler(c, info, resp)
+		} else {
+			usage, err = OaiResponsesHandler(c, info, resp)
+		}
+	case relayconstant.RelayModeResponsesCompact:
+		usage, err = OaiResponsesCompactionHandler(c, resp)
+	default:
+		if info.IsStream {
+			usage, err = OaiStreamHandler(c, info, resp)
+		} else {
+			usage, err = OpenaiHandler(c, info, resp)
+		}
+	}
+	return
+}
+
+func (a *Adaptor) GetModelList() []string {
+	switch a.ChannelType {
+	case constant.ChannelType360:
+		return ai360.ModelList
+	case constant.ChannelTypeLingYiWanWu:
+		return lingyiwanwu.ModelList
+	//case constant.ChannelTypeMiniMax:
+	//	return minimax.ModelList
+	case constant.ChannelTypeXinference:
+		return xinference.ModelList
+	case constant.ChannelTypeOpenRouter:
+		return openrouter.ModelList
+	default:
+		return ModelList
+	}
+}
+
+func (a *Adaptor) GetChannelName() string {
+	switch a.ChannelType {
+	case constant.ChannelType360:
+		return ai360.ChannelName
+	case constant.ChannelTypeLingYiWanWu:
+		return lingyiwanwu.ChannelName
+	//case constant.ChannelTypeMiniMax:
+	//	return minimax.ChannelName
+	case constant.ChannelTypeXinference:
+		return xinference.ChannelName
+	case constant.ChannelTypeOpenRouter:
+		return openrouter.ChannelName
+	default:
+		return ChannelName
+	}
+}
@@ -0,0 +1,149 @@
+package openai
+
+import (
+	"bytes"
+	"fmt"
+	"io"
+	"math"
+	"net/http"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	"github.com/QuantumNous/new-api/logger"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/QuantumNous/new-api/relay/helper"
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/QuantumNous/new-api/relaykit/types"
+	"github.com/QuantumNous/new-api/service"
+	"github.com/gin-gonic/gin"
+)
+
+func OpenaiTTSHandler(c *gin.Context, resp *http.Response, info *relaycommon.RelayInfo) *dto.Usage {
+	// the status code has been judged before, if there is a body reading failure,
+	// it should be regarded as a non-recoverable error, so it should not return err for external retry.
+	// Analogous to nginx's load balancing, it will only retry if it can't be requested or
+	// if the upstream returns a specific status code, once the upstream has already written the header,
+	// the subsequent failure of the response body should be regarded as a non-recoverable error,
+	// and can be terminated directly.
+	defer service.CloseResponseBodyGracefully(resp)
+	usage := &dto.Usage{}
+	usage.PromptTokens = info.GetEstimatePromptTokens()
+	usage.TotalTokens = info.GetEstimatePromptTokens()
+	for k, v := range resp.Header {
+		if !service.ShouldCopyUpstreamHeader(c, k, v) {
+			continue
+		}
+		c.Writer.Header().Set(k, v[0])
+	}
+	c.Writer.WriteHeader(resp.StatusCode)
+
+	if info.IsStream {
+		helper.StreamScannerHandler(c, resp, info, func(data string, sr *helper.StreamResult) {
+			if service.SundaySearch(data, "usage") {
+				var simpleResponse dto.SimpleResponse
+				if err := common.Unmarshal([]byte(data), &simpleResponse); err != nil {
+					logger.LogError(c, err.Error())
+					sr.Error(err)
+				} else if simpleResponse.Usage.TotalTokens != 0 {
+					usage.PromptTokens = simpleResponse.Usage.InputTokens
+					usage.CompletionTokens = simpleResponse.OutputTokens
+					usage.TotalTokens = simpleResponse.TotalTokens
+				}
+			}
+			if err := helper.StringData(c, data); err != nil {
+				sr.Error(err)
+			}
+		})
+	} else {
+		common.SetContextKey(c, constant.ContextKeyLocalCountTokens, true)
+		// 读取响应体到缓冲区
+		bodyBytes, err := io.ReadAll(resp.Body)
+		if err != nil {
+			logger.LogError(c, fmt.Sprintf("failed to read TTS response body: %v", err))
+			c.Writer.WriteHeaderNow()
+			return usage
+		}
+
+		// 写入响应到客户端
+		c.Writer.WriteHeaderNow()
+		_, err = c.Writer.Write(bodyBytes)
+		if err != nil {
+			logger.LogError(c, fmt.Sprintf("failed to write TTS response: %v", err))
+		}
+
+		// 计算音频时长并更新 usage
+		audioFormat := "mp3" // 默认格式
+		if audioReq, ok := info.Request.(*dto.AudioRequest); ok && audioReq.ResponseFormat != "" {
+			audioFormat = audioReq.ResponseFormat
+		}
+
+		var duration float64
+		var durationErr error
+
+		if audioFormat == "pcm" {
+			// PCM 格式没有文件头，根据 OpenAI TTS 的 PCM 参数计算时长
+			// 采样率: 24000 Hz, 位深度: 16-bit (2 bytes), 声道数: 1
+			const sampleRate = 24000
+			const bytesPerSample = 2
+			const channels = 1
+			duration = float64(len(bodyBytes)) / float64(sampleRate*bytesPerSample*channels)
+		} else {
+			ext := "." + audioFormat
+			reader := bytes.NewReader(bodyBytes)
+			duration, durationErr = common.GetAudioDuration(c.Request.Context(), reader, ext)
+		}
+
+		usage.PromptTokensDetails.TextTokens = usage.PromptTokens
+
+		if durationErr != nil {
+			logger.LogWarn(c, fmt.Sprintf("failed to get audio duration: %v", durationErr))
+			// 如果无法获取时长，则设置保底的 CompletionTokens，根据body大小计算
+			sizeInKB := float64(len(bodyBytes)) / 1000.0
+			estimatedTokens := int(math.Ceil(sizeInKB)) // 粗略估算每KB约等于1 token
+			usage.CompletionTokens = estimatedTokens
+			usage.CompletionTokenDetails.AudioTokens = estimatedTokens
+		} else if duration > 0 {
+			// 计算 token: ceil(duration) / 60.0 * 1000，即每分钟 1000 tokens。
+			// duration 解析自上游返回的音频元数据，饱和转换防止 int 回绕。
+			completionTokens := common.QuotaRound(math.Ceil(duration) / 60.0 * 1000)
+			usage.CompletionTokens = completionTokens
+			usage.CompletionTokenDetails.AudioTokens = completionTokens
+		}
+		usage.TotalTokens = usage.PromptTokens + usage.CompletionTokens
+	}
+
+	return usage
+}
+
+func OpenaiSTTHandler(c *gin.Context, resp *http.Response, info *relaycommon.RelayInfo, responseFormat string) (*types.NewAPIError, *dto.Usage) {
+	defer service.CloseResponseBodyGracefully(resp)
+
+	responseBody, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return types.NewOpenAIError(err, types.ErrorCodeReadResponseBodyFailed, http.StatusInternalServerError), nil
+	}
+	// 写入新的 response body
+	service.IOCopyBytesGracefully(c, resp, responseBody)
+
+	var responseData struct {
+		Usage *dto.Usage `json:"usage"`
+	}
+	if err := common.Unmarshal(responseBody, &responseData); err == nil && responseData.Usage != nil {
+		if responseData.Usage.TotalTokens > 0 {
+			usage := responseData.Usage
+			if usage.PromptTokens == 0 {
+				usage.PromptTokens = usage.InputTokens
+			}
+			if usage.CompletionTokens == 0 {
+				usage.CompletionTokens = usage.OutputTokens
+			}
+			return nil, usage
+		}
+	}
+
+	usage := &dto.Usage{}
+	usage.PromptTokens = info.GetEstimatePromptTokens()
+	usage.CompletionTokens = 0
+	usage.TotalTokens = usage.PromptTokens + usage.CompletionTokens
+	return nil, usage
+}
@@ -0,0 +1,342 @@
+package openai
+
+import (
+	"bufio"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+	"time"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/logger"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/QuantumNous/new-api/relay/helper"
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/QuantumNous/new-api/relaykit/relayconvert"
+	"github.com/QuantumNous/new-api/relaykit/types"
+	"github.com/QuantumNous/new-api/service"
+
+	"github.com/gin-gonic/gin"
+)
+
+func OaiResponsesToChatHandler(c *gin.Context, info *relaycommon.RelayInfo, resp *http.Response) (*dto.Usage, *types.NewAPIError) {
+	if resp == nil || resp.Body == nil {
+		return nil, types.NewOpenAIError(fmt.Errorf("invalid response"), types.ErrorCodeBadResponse, http.StatusInternalServerError)
+	}
+
+	defer service.CloseResponseBodyGracefully(resp)
+
+	var responsesResp dto.OpenAIResponsesResponse
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeReadResponseBodyFailed, http.StatusInternalServerError)
+	}
+
+	if err := common.Unmarshal(body, &responsesResp); err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+	}
+
+	if oaiError := responsesResp.GetOpenAIError(); oaiError != nil && oaiError.Type != "" {
+		return nil, types.WithOpenAIError(*oaiError, resp.StatusCode)
+	}
+
+	chatResult, err := relayconvert.ConvertResponse(c, info, types.RelayFormatOpenAI, &responsesResp)
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+	}
+	chatResp, ok := chatResult.Value.(*dto.OpenAITextResponse)
+	if !ok {
+		return nil, types.NewOpenAIError(fmt.Errorf("expected OpenAI chat response, got %T", chatResult.Value), types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+	}
+	if chatID := helper.GetResponseID(c); chatID != "" {
+		chatResp.Id = chatID
+	}
+	usage := chatResult.Usage
+
+	if usage == nil || usage.TotalTokens == 0 {
+		text := service.ExtractOutputTextFromResponses(&responsesResp)
+		usage = service.ResponseText2Usage(c, text, info.UpstreamModelName, info.GetEstimatePromptTokens())
+		chatResp.Usage = *usage
+	}
+
+	responseValue := any(chatResp)
+	if info.RelayFormat != types.RelayFormatOpenAI {
+		targetResult, err := relayconvert.ConvertResponse(c, info, info.RelayFormat, chatResp)
+		if err != nil {
+			return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+		}
+		responseValue = targetResult.Value
+	}
+	responseBody, err := common.Marshal(responseValue)
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeJsonMarshalFailed, http.StatusInternalServerError)
+	}
+
+	service.IOCopyBytesGracefully(c, resp, responseBody)
+	return usage, nil
+}
+
+func OaiResponsesToChatBufferedStreamHandler(c *gin.Context, info *relaycommon.RelayInfo, resp *http.Response) (*dto.Usage, *types.NewAPIError) {
+	if resp == nil || resp.Body == nil {
+		return nil, types.NewOpenAIError(fmt.Errorf("invalid response"), types.ErrorCodeBadResponse, http.StatusInternalServerError)
+	}
+	defer service.CloseResponseBodyGracefully(resp)
+
+	accumulator := relayconvert.NewResponsesBufferedAccumulator()
+	var finalResponse *dto.OpenAIResponsesResponse
+	var streamErr *types.NewAPIError
+
+	scanner := helper.NewStreamScanner(resp.Body)
+	scanner.Split(bufio.ScanLines)
+	for scanner.Scan() {
+		line := scanner.Text()
+		if len(line) < 6 || line[:5] != "data:" {
+			continue
+		}
+		data := line[5:]
+		data = strings.TrimSpace(data)
+		if data == "" || data == "[DONE]" {
+			if data == "[DONE]" {
+				break
+			}
+			continue
+		}
+
+		var streamResp dto.ResponsesStreamResponse
+		if err := common.UnmarshalJsonStr(data, &streamResp); err != nil {
+			logger.LogError(c, "failed to unmarshal buffered responses stream event: "+err.Error())
+			streamErr = types.NewOpenAIError(err, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+			break
+		}
+		accumulator.ProcessEvent(&streamResp)
+		switch streamResp.Type {
+		case "response.completed", "response.done", "response.incomplete":
+			finalResponse = streamResp.Response
+			if streamResp.Type == "response.incomplete" {
+				if finalResponse == nil {
+					finalResponse = &dto.OpenAIResponsesResponse{}
+				}
+				if len(finalResponse.Status) == 0 {
+					finalResponse.Status = []byte(`"incomplete"`)
+				}
+			}
+		case "response.failed", "response.error":
+			if streamResp.Response != nil {
+				if oaiErr := streamResp.Response.GetOpenAIError(); oaiErr != nil && oaiErr.Type != "" {
+					streamErr = types.WithOpenAIError(*oaiErr, http.StatusInternalServerError)
+					break
+				}
+			}
+			streamErr = types.NewOpenAIError(fmt.Errorf("responses stream error: %s", streamResp.Type), types.ErrorCodeBadResponse, http.StatusInternalServerError)
+		}
+		if streamErr != nil || finalResponse != nil {
+			break
+		}
+	}
+	if streamErr != nil {
+		return nil, streamErr
+	}
+	if err := scanner.Err(); err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponse, http.StatusInternalServerError)
+	}
+	if finalResponse == nil {
+		finalResponse = &dto.OpenAIResponsesResponse{
+			ID:        helper.GetResponseID(c),
+			CreatedAt: int(time.Now().Unix()),
+			Model:     info.UpstreamModelName,
+			Status:    []byte(`"completed"`),
+		}
+	}
+	accumulator.SupplementResponseOutput(finalResponse)
+
+	chatResult, err := relayconvert.ConvertResponse(c, info, types.RelayFormatOpenAI, finalResponse)
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+	}
+	chatResp, ok := chatResult.Value.(*dto.OpenAITextResponse)
+	if !ok {
+		return nil, types.NewOpenAIError(fmt.Errorf("expected OpenAI chat response, got %T", chatResult.Value), types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+	}
+	if chatID := helper.GetResponseID(c); chatID != "" {
+		chatResp.Id = chatID
+	}
+	usage := chatResult.Usage
+	if usage == nil || usage.TotalTokens == 0 {
+		text := service.ExtractOutputTextFromResponses(finalResponse)
+		usage = service.ResponseText2Usage(c, text, info.UpstreamModelName, info.GetEstimatePromptTokens())
+		chatResp.Usage = *usage
+	}
+
+	responseValue := any(chatResp)
+	if info.RelayFormat != types.RelayFormatOpenAI {
+		targetResult, err := relayconvert.ConvertResponse(c, info, info.RelayFormat, chatResp)
+		if err != nil {
+			return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+		}
+		responseValue = targetResult.Value
+	}
+	responseBody, err := common.Marshal(responseValue)
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeJsonMarshalFailed, http.StatusInternalServerError)
+	}
+
+	service.IOCopyBytesGracefully(c, resp, responseBody)
+	return usage, nil
+}
+
+func OaiResponsesToChatStreamHandler(c *gin.Context, info *relaycommon.RelayInfo, resp *http.Response) (*dto.Usage, *types.NewAPIError) {
+	if resp == nil || resp.Body == nil {
+		return nil, types.NewOpenAIError(fmt.Errorf("invalid response"), types.ErrorCodeBadResponse, http.StatusInternalServerError)
+	}
+
+	defer service.CloseResponseBodyGracefully(resp)
+
+	responseId := helper.GetResponseID(c)
+	createAt := time.Now().Unix()
+	state, err := relayconvert.NewResponseStreamState(types.RelayFormatOpenAIResponses, info.RelayFormat, relayconvert.ResponseStreamOptions{
+		ID:      responseId,
+		Model:   info.UpstreamModelName,
+		Created: createAt,
+	})
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponse, http.StatusInternalServerError)
+	}
+	streamErr := (*types.NewAPIError)(nil)
+
+	if info.RelayFormat == types.RelayFormatClaude && info.ClaudeConvertInfo == nil {
+		info.ClaudeConvertInfo = &relaycommon.ClaudeConvertInfo{LastMessagesType: relaycommon.LastMessageTypeNone}
+	}
+
+	sendGeminiResponse := func(geminiResponse *dto.GeminiChatResponse) bool {
+		if geminiResponse == nil {
+			return true
+		}
+		geminiResponseStr, err := common.Marshal(geminiResponse)
+		if err != nil {
+			streamErr = types.NewOpenAIError(err, types.ErrorCodeJsonMarshalFailed, http.StatusInternalServerError)
+			return false
+		}
+		c.Render(-1, common.CustomEvent{Data: "data: " + string(geminiResponseStr)})
+		_ = helper.FlushWriter(c)
+		return true
+	}
+
+	sendStreamResult := func(result relayconvert.ResponseResult) bool {
+		switch value := result.Value.(type) {
+		case dto.ChatCompletionsStreamResponse:
+			if len(value.Choices) == 0 && value.Usage == nil {
+				return true
+			}
+			if err := helper.ObjectData(c, &value); err != nil {
+				streamErr = types.NewOpenAIError(err, types.ErrorCodeBadResponse, http.StatusInternalServerError)
+				return false
+			}
+			return true
+		case *dto.ChatCompletionsStreamResponse:
+			if value == nil || (len(value.Choices) == 0 && value.Usage == nil) {
+				return true
+			}
+			if err := helper.ObjectData(c, value); err != nil {
+				streamErr = types.NewOpenAIError(err, types.ErrorCodeBadResponse, http.StatusInternalServerError)
+				return false
+			}
+			return true
+		case dto.ClaudeResponse:
+			if err := helper.ClaudeData(c, value); err != nil {
+				streamErr = types.NewOpenAIError(err, types.ErrorCodeBadResponse, http.StatusInternalServerError)
+				return false
+			}
+			return true
+		case *dto.ClaudeResponse:
+			if value == nil {
+				return true
+			}
+			if err := helper.ClaudeData(c, *value); err != nil {
+				streamErr = types.NewOpenAIError(err, types.ErrorCodeBadResponse, http.StatusInternalServerError)
+				return false
+			}
+			return true
+		case dto.GeminiChatResponse:
+			return sendGeminiResponse(&value)
+		case *dto.GeminiChatResponse:
+			return sendGeminiResponse(value)
+		default:
+			streamErr = types.NewOpenAIError(fmt.Errorf("unsupported converted stream response type %T", result.Value), types.ErrorCodeBadResponse, http.StatusInternalServerError)
+			return false
+		}
+	}
+
+	helper.StreamScannerHandler(c, resp, info, func(data string, sr *helper.StreamResult) {
+		if streamErr != nil {
+			sr.Stop(streamErr)
+			return
+		}
+
+		var streamResp dto.ResponsesStreamResponse
+		if err := common.UnmarshalJsonStr(data, &streamResp); err != nil {
+			logger.LogError(c, "failed to unmarshal responses stream event: "+err.Error())
+			sr.Error(err)
+			return
+		}
+
+		if streamResp.Type == "response.error" || streamResp.Type == "response.failed" {
+			if streamResp.Response != nil {
+				if oaiErr := streamResp.Response.GetOpenAIError(); oaiErr != nil && oaiErr.Type != "" {
+					streamErr = types.WithOpenAIError(*oaiErr, http.StatusInternalServerError)
+					sr.Stop(streamErr)
+					return
+				}
+			}
+			streamErr = types.NewOpenAIError(fmt.Errorf("responses stream error: %s", streamResp.Type), types.ErrorCodeBadResponse, http.StatusInternalServerError)
+			sr.Stop(streamErr)
+			return
+		}
+
+		results, err := relayconvert.ConvertStreamResponseChunk(c, info, state, &streamResp)
+		if err != nil {
+			streamErr = types.NewOpenAIError(err, types.ErrorCodeBadResponse, http.StatusInternalServerError)
+			sr.Stop(streamErr)
+			return
+		}
+		for _, result := range results {
+			if !sendStreamResult(result) {
+				sr.Stop(streamErr)
+				return
+			}
+		}
+	})
+
+	if streamErr != nil {
+		return nil, streamErr
+	}
+
+	usage := state.Usage()
+	if usage == nil || usage.TotalTokens == 0 {
+		usage = service.ResponseText2Usage(c, state.UsageText(), info.UpstreamModelName, info.GetEstimatePromptTokens())
+		state.SetUsage(usage)
+	}
+
+	if info.RelayFormat == types.RelayFormatClaude && info.ClaudeConvertInfo != nil {
+		info.ClaudeConvertInfo.Usage = usage
+	}
+	finalResults, err := relayconvert.FinalizeStreamResponse(c, info, state)
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponse, http.StatusInternalServerError)
+	}
+	for _, result := range finalResults {
+		if !sendStreamResult(result) {
+			return nil, streamErr
+		}
+	}
+	if info.RelayFormat == types.RelayFormatOpenAI && info.ShouldIncludeUsage && usage != nil {
+		if err := helper.ObjectData(c, helper.GenerateFinalUsageResponse(responseId, createAt, info.UpstreamModelName, *usage)); err != nil {
+			return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponse, http.StatusInternalServerError)
+		}
+	}
+
+	if info.RelayFormat == types.RelayFormatOpenAI {
+		helper.Done(c)
+	}
+	return usage, nil
+}
@@ -0,0 +1,235 @@
+package openai
+
+import (
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/QuantumNous/new-api/relaykit/types"
+	"github.com/gin-gonic/gin"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func newResponsesChatTestContext(t *testing.T, body string, isStream bool) (*gin.Context, *httptest.ResponseRecorder, *http.Response, *relaycommon.RelayInfo) {
+	t.Helper()
+
+	recorder := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(recorder)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
+	c.Set(common.RequestIdKey, "responses-test")
+
+	resp := &http.Response{
+		StatusCode: http.StatusOK,
+		Body:       io.NopCloser(strings.NewReader(body)),
+		Header:     http.Header{"Content-Type": []string{"text/event-stream"}},
+	}
+	info := &relaycommon.RelayInfo{
+		ChannelMeta:        &relaycommon.ChannelMeta{UpstreamModelName: "gpt-test"},
+		IsStream:           isStream,
+		RelayFormat:        types.RelayFormatOpenAI,
+		ShouldIncludeUsage: true,
+		DisablePing:        true,
+	}
+	return c, recorder, resp, info
+}
+
+func TestOaiResponsesToChatStreamHandlerConvertsSSEOrderAndUsage(t *testing.T) {
+	oldMode := gin.Mode()
+	gin.SetMode(gin.TestMode)
+	t.Cleanup(func() { gin.SetMode(oldMode) })
+
+	oldTimeout := constant.StreamingTimeout
+	constant.StreamingTimeout = 30
+	t.Cleanup(func() { constant.StreamingTimeout = oldTimeout })
+
+	body := strings.Join([]string{
+		`data: {"type":"response.created","response":{"id":"resp_1","model":"gpt-test","created_at":1710000000}}`,
+		`data: {"type":"response.output_text.delta","delta":"hello"}`,
+		`data: {"type":"response.output_item.added","output_index":1,"item":{"type":"function_call","id":"fc_1","call_id":"call_1","name":"lookup"}}`,
+		`data: {"type":"response.function_call_arguments.delta","output_index":1,"delta":"{\"q\":\"x\"}"}`,
+		`data: {"type":"response.completed","response":{"status":"completed","usage":{"input_tokens":2,"output_tokens":3,"total_tokens":5}}}`,
+		`data: [DONE]`,
+		``,
+	}, "\n")
+
+	c, recorder, resp, info := newResponsesChatTestContext(t, body, true)
+
+	usage, err := OaiResponsesToChatStreamHandler(c, info, resp)
+	require.Nil(t, err)
+	require.NotNil(t, usage)
+	require.Equal(t, 2, usage.PromptTokens)
+	require.Equal(t, 3, usage.CompletionTokens)
+	require.Equal(t, 5, usage.TotalTokens)
+
+	got := recorder.Body.String()
+	require.Equal(t, "text/event-stream", recorder.Header().Get("Content-Type"))
+	require.Contains(t, got, `"role":"assistant"`)
+	require.Contains(t, got, `"content":"hello"`)
+	require.Contains(t, got, `"name":"lookup"`)
+	require.Contains(t, got, `"arguments":"{\"q\":\"x\"}"`)
+	require.Contains(t, got, `"finish_reason":"tool_calls"`)
+	require.Contains(t, got, `"usage":{"prompt_tokens":2,"completion_tokens":3,"total_tokens":5`)
+	require.Contains(t, got, `data: [DONE]`)
+	requireOrderedSubstrings(t, got,
+		`"role":"assistant"`,
+		`"content":"hello"`,
+		`"name":"lookup"`,
+		`"arguments":"{\"q\":\"x\"}"`,
+		`"finish_reason":"tool_calls"`,
+		`"usage":{"prompt_tokens":2,"completion_tokens":3,"total_tokens":5`,
+		`data: [DONE]`,
+	)
+}
+
+func TestOaiResponsesToChatStreamHandlerConvertsClaudeSSETerminalsAndUsage(t *testing.T) {
+	oldMode := gin.Mode()
+	gin.SetMode(gin.TestMode)
+	t.Cleanup(func() { gin.SetMode(oldMode) })
+
+	oldTimeout := constant.StreamingTimeout
+	constant.StreamingTimeout = 30
+	t.Cleanup(func() { constant.StreamingTimeout = oldTimeout })
+
+	body := strings.Join([]string{
+		`data: {"type":"response.created","response":{"id":"resp_1","model":"gpt-test","created_at":1710000000}}`,
+		`data: {"type":"response.output_text.delta","delta":"hello"}`,
+		`data: {"type":"response.completed","response":{"status":"completed","usage":{"input_tokens":2,"output_tokens":3,"total_tokens":5}}}`,
+		`data: [DONE]`,
+		``,
+	}, "\n")
+
+	c, recorder, resp, info := newResponsesChatTestContext(t, body, true)
+	info.RelayFormat = types.RelayFormatClaude
+
+	usage, err := OaiResponsesToChatStreamHandler(c, info, resp)
+	require.Nil(t, err)
+	require.NotNil(t, usage)
+	assert.Equal(t, 2, usage.PromptTokens)
+	assert.Equal(t, 3, usage.CompletionTokens)
+	assert.Equal(t, 5, usage.TotalTokens)
+
+	got := recorder.Body.String()
+	assert.Equal(t, "text/event-stream", recorder.Header().Get("Content-Type"))
+	assert.Equal(t, 1, strings.Count(got, "event: message_start\n"))
+	assert.Equal(t, 1, strings.Count(got, "event: content_block_stop\n"))
+	assert.Equal(t, 1, strings.Count(got, "event: message_delta\n"))
+	assert.Equal(t, 1, strings.Count(got, "event: message_stop\n"))
+
+	messageDeltaFrame := ""
+	for _, frame := range strings.Split(got, "\n\n") {
+		if strings.HasPrefix(frame, "event: message_delta\n") {
+			messageDeltaFrame = frame
+			break
+		}
+	}
+	require.NotEmpty(t, messageDeltaFrame)
+	assert.Contains(t, messageDeltaFrame, `"type":"message_delta"`)
+	assert.Contains(t, messageDeltaFrame, `"stop_reason":"end_turn"`)
+	assert.Contains(t, messageDeltaFrame, `"input_tokens":2`)
+	assert.Contains(t, messageDeltaFrame, `"output_tokens":3`)
+	requireOrderedSubstrings(t, got,
+		"event: message_start\n",
+		"event: content_block_stop\n",
+		"event: message_delta\n",
+		"event: message_stop\n",
+	)
+}
+
+func TestOaiResponsesToChatBufferedStreamHandlerReturnsJSONFromSSE(t *testing.T) {
+	oldMode := gin.Mode()
+	gin.SetMode(gin.TestMode)
+	t.Cleanup(func() { gin.SetMode(oldMode) })
+
+	body := strings.Join([]string{
+		`data: {"type":"response.output_text.delta","delta":"buffered text"}`,
+		`data: {"type":"response.output_item.added","output_index":0,"item":{"type":"function_call","id":"fc_1","call_id":"call_1","name":"lookup"}}`,
+		`data: {"type":"response.function_call_arguments.delta","output_index":0,"delta":"{\"q\":\"x\"}"}`,
+		`data: {"type":"response.done","response":{"model":"gpt-test","status":"completed","usage":{"input_tokens":1,"output_tokens":2,"total_tokens":3}}}`,
+		`data: [DONE]`,
+		``,
+	}, "\n")
+
+	c, recorder, resp, info := newResponsesChatTestContext(t, body, false)
+
+	usage, err := OaiResponsesToChatBufferedStreamHandler(c, info, resp)
+	require.Nil(t, err)
+	require.NotNil(t, usage)
+	require.Equal(t, 3, usage.TotalTokens)
+
+	got := recorder.Body.String()
+	require.NotContains(t, got, `data:`)
+	require.Contains(t, got, `"object":"chat.completion"`)
+	require.Contains(t, got, `"content":"buffered text"`)
+	require.Contains(t, got, `"name":"lookup"`)
+	require.Contains(t, got, `"arguments":"{\"q\":\"x\"}"`)
+	require.Contains(t, got, `"finish_reason":"tool_calls"`)
+}
+
+func TestOaiChatToResponsesStreamHandlerConvertsSSEOrderAndUsage(t *testing.T) {
+	oldMode := gin.Mode()
+	gin.SetMode(gin.TestMode)
+	t.Cleanup(func() { gin.SetMode(oldMode) })
+
+	oldTimeout := constant.StreamingTimeout
+	constant.StreamingTimeout = 30
+	t.Cleanup(func() { constant.StreamingTimeout = oldTimeout })
+
+	body := strings.Join([]string{
+		`data: {"id":"chatcmpl_1","object":"chat.completion.chunk","created":1710000000,"model":"gpt-test","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}`,
+		`data: {"id":"chatcmpl_1","object":"chat.completion.chunk","created":1710000000,"model":"gpt-test","choices":[{"index":0,"delta":{"content":"hello"},"finish_reason":null}]}`,
+		`data: {"id":"chatcmpl_1","object":"chat.completion.chunk","created":1710000000,"model":"gpt-test","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"id":"call_1","type":"function","function":{"name":"lookup"}}]},"finish_reason":null}]}`,
+		`data: {"id":"chatcmpl_1","object":"chat.completion.chunk","created":1710000000,"model":"gpt-test","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"{\"q\":\"x\"}"}}]},"finish_reason":null}]}`,
+		`data: {"id":"chatcmpl_1","object":"chat.completion.chunk","created":1710000000,"model":"gpt-test","choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}]}`,
+		`data: {"id":"chatcmpl_1","object":"chat.completion.chunk","created":1710000000,"model":"gpt-test","choices":[],"usage":{"prompt_tokens":2,"completion_tokens":3,"total_tokens":5}}`,
+		`data: [DONE]`,
+		``,
+	}, "\n")
+
+	c, recorder, resp, info := newResponsesChatTestContext(t, body, true)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
+
+	usage, err := OaiChatToResponsesStreamHandler(c, info, resp)
+	require.Nil(t, err)
+	require.NotNil(t, usage)
+	require.Equal(t, 2, usage.PromptTokens)
+	require.Equal(t, 3, usage.CompletionTokens)
+	require.Equal(t, 5, usage.TotalTokens)
+
+	got := recorder.Body.String()
+	require.Equal(t, "text/event-stream", recorder.Header().Get("Content-Type"))
+	require.Contains(t, got, `event: response.created`)
+	require.Contains(t, got, `event: response.output_text.delta`)
+	require.Contains(t, got, `"delta":"hello"`)
+	require.Contains(t, got, `event: response.function_call_arguments.delta`)
+	require.Contains(t, got, `"delta":"{\"q\":\"x\"}"`)
+	require.Contains(t, got, `event: response.completed`)
+	require.Contains(t, got, `"input_tokens":2`)
+	require.Contains(t, got, `"output_tokens":3`)
+	requireOrderedSubstrings(t, got,
+		`event: response.created`,
+		`event: response.output_item.added`,
+		`event: response.output_text.delta`,
+		`event: response.output_item.added`,
+		`event: response.function_call_arguments.delta`,
+		`event: response.output_text.done`,
+		`event: response.function_call_arguments.done`,
+		`event: response.completed`,
+	)
+}
+
+func requireOrderedSubstrings(t *testing.T, s string, parts ...string) {
+	t.Helper()
+
+	offset := 0
+	for _, part := range parts {
+		idx := strings.Index(s[offset:], part)
+		require.NotEqualf(t, -1, idx, "missing %q after byte offset %d", part, offset)
+		offset += idx + len(part)
+	}
+}
@@ -0,0 +1,78 @@
+package openai
+
+var ModelList = []string{
+	"gpt-3.5-turbo", "gpt-3.5-turbo-0613", "gpt-3.5-turbo-1106", "gpt-3.5-turbo-0125",
+	"gpt-3.5-turbo-16k", "gpt-3.5-turbo-16k-0613",
+	"gpt-3.5-turbo-instruct", "gpt-3.5-turbo-instruct-0914",
+	"gpt-4", "gpt-4-0613", "gpt-4-1106-preview", "gpt-4-0125-preview",
+	"gpt-4-32k", "gpt-4-32k-0613",
+	"gpt-4-turbo-preview", "gpt-4-turbo", "gpt-4-turbo-2024-04-09",
+	"gpt-4-vision-preview",
+	"chatgpt-4o-latest",
+	"gpt-4o", "gpt-4o-2024-05-13", "gpt-4o-2024-08-06", "gpt-4o-2024-11-20",
+	"gpt-4o-transcribe", "gpt-4o-transcribe-diarize",
+	"gpt-4o-search-preview", "gpt-4o-search-preview-2025-03-11",
+	"gpt-4o-mini", "gpt-4o-mini-2024-07-18",
+	"gpt-4o-mini-transcribe", "gpt-4o-mini-transcribe-2025-03-20", "gpt-4o-mini-transcribe-2025-12-15",
+	"gpt-4o-mini-tts", "gpt-4o-mini-tts-2025-03-20", "gpt-4o-mini-tts-2025-12-15",
+	"gpt-4o-mini-search-preview", "gpt-4o-mini-search-preview-2025-03-11",
+	"gpt-4.5-preview", "gpt-4.5-preview-2025-02-27",
+	"gpt-4.1", "gpt-4.1-2025-04-14",
+	"gpt-4.1-mini", "gpt-4.1-mini-2025-04-14",
+	"gpt-4.1-nano", "gpt-4.1-nano-2025-04-14",
+	"o1", "o1-2024-12-17",
+	"o1-preview", "o1-preview-2024-09-12",
+	"o1-mini", "o1-mini-2024-09-12",
+	"o1-pro", "o1-pro-2025-03-19",
+	"o3-mini", "o3-mini-2025-01-31",
+	"o3-mini-high", "o3-mini-2025-01-31-high",
+	"o3-mini-low", "o3-mini-2025-01-31-low",
+	"o3-mini-medium", "o3-mini-2025-01-31-medium",
+	"o3", "o3-2025-04-16",
+	"o3-pro", "o3-pro-2025-06-10",
+	"o3-deep-research", "o3-deep-research-2025-06-26",
+	"o4-mini", "o4-mini-2025-04-16",
+	"o4-mini-deep-research", "o4-mini-deep-research-2025-06-26",
+	"gpt-5", "gpt-5-2025-08-07", "gpt-5-chat-latest",
+	"gpt-5-mini", "gpt-5-mini-2025-08-07",
+	"gpt-5-nano", "gpt-5-nano-2025-08-07",
+	"gpt-5-codex",
+	"gpt-5-pro", "gpt-5-pro-2025-10-06",
+	"gpt-5-search-api", "gpt-5-search-api-2025-10-14",
+	"gpt-5.1", "gpt-5.1-2025-11-13", "gpt-5.1-chat-latest",
+	"gpt-5.1-codex", "gpt-5.1-codex-mini", "gpt-5.1-codex-max",
+	"gpt-5.2", "gpt-5.2-2025-12-11", "gpt-5.2-chat-latest",
+	"gpt-5.2-pro", "gpt-5.2-pro-2025-12-11",
+	"gpt-5.2-codex",
+	"gpt-5.3-chat-latest",
+	"gpt-5.3-codex",
+	"gpt-5.4", "gpt-5.4-2026-03-05",
+	"gpt-5.4-pro", "gpt-5.4-pro-2026-03-05",
+	"gpt-4o-audio-preview", "gpt-4o-audio-preview-2024-10-01", "gpt-4o-audio-preview-2024-12-17", "gpt-4o-audio-preview-2025-06-03",
+	"gpt-4o-realtime-preview", "gpt-4o-realtime-preview-2024-10-01", "gpt-4o-realtime-preview-2024-12-17", "gpt-4o-realtime-preview-2025-06-03",
+	"gpt-4o-mini-realtime-preview", "gpt-4o-mini-realtime-preview-2024-12-17",
+	"gpt-4o-mini-audio-preview", "gpt-4o-mini-audio-preview-2024-12-17",
+	"gpt-audio", "gpt-audio-2025-08-28",
+	"gpt-audio-mini", "gpt-audio-mini-2025-10-06", "gpt-audio-mini-2025-12-15",
+	"gpt-audio-1.5",
+	"gpt-realtime", "gpt-realtime-2025-08-28",
+	"gpt-realtime-mini", "gpt-realtime-mini-2025-10-06", "gpt-realtime-mini-2025-12-15",
+	"gpt-realtime-1.5",
+	"gpt-realtime-2", "gpt-realtime-2.1", "gpt-realtime-2.1-mini",
+	"gpt-realtime-whisper", "gpt-realtime-translate",
+	"text-embedding-ada-002", "text-embedding-3-small", "text-embedding-3-large",
+	"text-curie-001", "text-babbage-001", "text-ada-001",
+	"text-moderation-latest", "text-moderation-stable",
+	"omni-moderation-latest", "omni-moderation-2024-09-26",
+	"text-davinci-edit-001",
+	"davinci-002", "babbage-002",
+	"dall-e-2", "dall-e-3",
+	"gpt-image-1", "gpt-image-1-mini", "gpt-image-1.5",
+	"chatgpt-image-latest",
+	"whisper-1",
+	"tts-1", "tts-1-1106", "tts-1-hd", "tts-1-hd-1106",
+	"computer-use-preview", "computer-use-preview-2025-03-11",
+	"sora-2", "sora-2-pro",
+}
+
+var ChannelName = "openai"
@@ -0,0 +1,244 @@
+package openai
+
+import (
+	"fmt"
+	"strings"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/logger"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	relayconstant "github.com/QuantumNous/new-api/relay/constant"
+	"github.com/QuantumNous/new-api/relay/helper"
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/QuantumNous/new-api/relaykit/relayconvert"
+	"github.com/QuantumNous/new-api/relaykit/types"
+	"github.com/QuantumNous/new-api/service"
+
+	"github.com/samber/lo"
+
+	"github.com/gin-gonic/gin"
+)
+
+// 辅助函数
+func HandleStreamFormat(c *gin.Context, info *relaycommon.RelayInfo, data string, forceFormat bool, thinkToContent bool) error {
+	info.SendResponseCount++
+
+	switch info.RelayFormat {
+	case types.RelayFormatOpenAI:
+		return sendStreamData(c, info, data, forceFormat, thinkToContent)
+	case types.RelayFormatClaude:
+		return handleClaudeFormat(c, data, info)
+	case types.RelayFormatGemini:
+		return handleGeminiFormat(c, data, info)
+	}
+	return nil
+}
+
+func handleClaudeFormat(c *gin.Context, data string, info *relaycommon.RelayInfo) error {
+	var streamResponse dto.ChatCompletionsStreamResponse
+	if err := common.Unmarshal(common.StringToByteSlice(data), &streamResponse); err != nil {
+		return err
+	}
+
+	if streamResponse.Usage != nil {
+		info.ClaudeConvertInfo.Usage = streamResponse.Usage
+	}
+	result, err := relayconvert.ConvertStreamResponse(c, info, types.RelayFormatClaude, &streamResponse)
+	if err != nil {
+		return err
+	}
+	claudeResponses, ok := result.Value.([]*dto.ClaudeResponse)
+	if !ok {
+		return fmt.Errorf("expected Claude stream responses, got %T", result.Value)
+	}
+	for _, resp := range claudeResponses {
+		helper.ClaudeData(c, *resp)
+	}
+	return nil
+}
+
+func handleGeminiFormat(c *gin.Context, data string, info *relaycommon.RelayInfo) error {
+	var streamResponse dto.ChatCompletionsStreamResponse
+	if err := common.Unmarshal(common.StringToByteSlice(data), &streamResponse); err != nil {
+		logger.LogError(c, "failed to unmarshal stream response: "+err.Error())
+		return err
+	}
+
+	result, err := relayconvert.ConvertStreamResponse(c, info, types.RelayFormatGemini, &streamResponse)
+	if err != nil {
+		return err
+	}
+	geminiResponse, ok := result.Value.(*dto.GeminiChatResponse)
+	if !ok {
+		return fmt.Errorf("expected Gemini stream response, got %T", result.Value)
+	}
+
+	// 如果返回 nil，表示没有实际内容，跳过发送
+	if geminiResponse == nil {
+		return nil
+	}
+
+	geminiResponseStr, err := common.Marshal(geminiResponse)
+	if err != nil {
+		logger.LogError(c, "failed to marshal gemini response: "+err.Error())
+		return err
+	}
+
+	// send gemini format response
+	c.Render(-1, common.CustomEvent{Data: "data: " + string(geminiResponseStr)})
+	_ = helper.FlushWriter(c)
+	return nil
+}
+
+func ProcessStreamResponse(streamResponse dto.ChatCompletionsStreamResponse, responseTextBuilder *strings.Builder, toolCount *int) error {
+	for _, choice := range streamResponse.Choices {
+		responseTextBuilder.WriteString(choice.Delta.GetContentString())
+		responseTextBuilder.WriteString(choice.Delta.GetReasoningContent())
+		if choice.Delta.ToolCalls != nil {
+			if len(choice.Delta.ToolCalls) > *toolCount {
+				*toolCount = len(choice.Delta.ToolCalls)
+			}
+			for _, tool := range choice.Delta.ToolCalls {
+				responseTextBuilder.WriteString(tool.Function.Name)
+				responseTextBuilder.WriteString(tool.Function.Arguments)
+			}
+		}
+	}
+	return nil
+}
+
+func processTokenData(relayMode int, data string, responseTextBuilder *strings.Builder, toolCount *int) error {
+	switch relayMode {
+	case relayconstant.RelayModeChatCompletions:
+		var streamResponse dto.ChatCompletionsStreamResponse
+		if err := common.UnmarshalJsonStr(data, &streamResponse); err != nil {
+			return err
+		}
+		return ProcessStreamResponse(streamResponse, responseTextBuilder, toolCount)
+	case relayconstant.RelayModeCompletions:
+		var streamResponse dto.CompletionsStreamResponse
+		if err := common.UnmarshalJsonStr(data, &streamResponse); err != nil {
+			return err
+		}
+		processCompletionsStreamResponse(streamResponse, responseTextBuilder)
+	}
+	return nil
+}
+
+func processCompletionsStreamResponse(streamResponse dto.CompletionsStreamResponse, responseTextBuilder *strings.Builder) {
+	for _, choice := range streamResponse.Choices {
+		responseTextBuilder.WriteString(choice.Text)
+	}
+}
+
+func handleLastResponse(lastStreamData string, responseId *string, createAt *int64,
+	systemFingerprint *string, model *string, usage **dto.Usage,
+	containStreamUsage *bool, info *relaycommon.RelayInfo,
+	shouldSendLastResp *bool) error {
+
+	var lastStreamResponse dto.ChatCompletionsStreamResponse
+	if err := common.Unmarshal(common.StringToByteSlice(lastStreamData), &lastStreamResponse); err != nil {
+		return err
+	}
+
+	*responseId = lastStreamResponse.Id
+	*createAt = lastStreamResponse.Created
+	*systemFingerprint = lastStreamResponse.GetSystemFingerprint()
+	*model = lastStreamResponse.Model
+
+	if service.ValidUsage(lastStreamResponse.Usage) {
+		*containStreamUsage = true
+		*usage = lastStreamResponse.Usage
+		if !info.ShouldIncludeUsage {
+			*shouldSendLastResp = lo.SomeBy(lastStreamResponse.Choices, func(choice dto.ChatCompletionsStreamResponseChoice) bool {
+				return choice.Delta.GetContentString() != "" || choice.Delta.GetReasoningContent() != ""
+			})
+		}
+	}
+
+	return nil
+}
+
+func HandleFinalResponse(c *gin.Context, info *relaycommon.RelayInfo, lastStreamData string,
+	responseId string, createAt int64, model string, systemFingerprint string,
+	usage *dto.Usage, containStreamUsage bool) {
+
+	switch info.RelayFormat {
+	case types.RelayFormatOpenAI:
+		if info.ShouldIncludeUsage && !containStreamUsage {
+			response := helper.GenerateFinalUsageResponse(responseId, createAt, model, *usage)
+			response.SetSystemFingerprint(systemFingerprint)
+			helper.ObjectData(c, response)
+		}
+		helper.Done(c)
+
+	case types.RelayFormatClaude:
+		var streamResponse dto.ChatCompletionsStreamResponse
+		if err := common.Unmarshal(common.StringToByteSlice(lastStreamData), &streamResponse); err != nil {
+			common.SysLog("error unmarshalling stream response: " + err.Error())
+			return
+		}
+
+		info.ClaudeConvertInfo.Usage = usage
+
+		result, err := relayconvert.ConvertStreamResponse(c, info, types.RelayFormatClaude, &streamResponse)
+		if err != nil {
+			common.SysLog("error converting Claude stream response: " + err.Error())
+			return
+		}
+		claudeResponses, ok := result.Value.([]*dto.ClaudeResponse)
+		if !ok {
+			common.SysLog(fmt.Sprintf("expected Claude stream responses, got %T", result.Value))
+			return
+		}
+		for _, resp := range claudeResponses {
+			_ = helper.ClaudeData(c, *resp)
+		}
+		info.ClaudeConvertInfo.Done = true
+
+	case types.RelayFormatGemini:
+		var streamResponse dto.ChatCompletionsStreamResponse
+		if err := common.Unmarshal(common.StringToByteSlice(lastStreamData), &streamResponse); err != nil {
+			common.SysLog("error unmarshalling stream response: " + err.Error())
+			return
+		}
+
+		// 这里处理的是 openai 最后一个流响应，其 delta 为空，有 finish_reason 字段
+		// 因此相比较于 google 官方的流响应，由 openai 转换而来会多一个 parts 为空，finishReason 为 STOP 的响应
+		// 而包含最后一段文本输出的响应（倒数第二个）的 finishReason 为 null
+		// 暂不知是否有程序会不兼容。
+
+		result, err := relayconvert.ConvertStreamResponse(c, info, types.RelayFormatGemini, &streamResponse)
+		if err != nil {
+			common.SysLog("error converting Gemini stream response: " + err.Error())
+			return
+		}
+		geminiResponse, ok := result.Value.(*dto.GeminiChatResponse)
+		if !ok {
+			common.SysLog(fmt.Sprintf("expected Gemini stream response, got %T", result.Value))
+			return
+		}
+
+		// openai 流响应开头的空数据
+		if geminiResponse == nil {
+			return
+		}
+
+		geminiResponseStr, err := common.Marshal(geminiResponse)
+		if err != nil {
+			common.SysLog("error marshalling gemini response: " + err.Error())
+			return
+		}
+
+		// 发送最终的 Gemini 响应
+		c.Render(-1, common.CustomEvent{Data: "data: " + string(geminiResponseStr)})
+		_ = helper.FlushWriter(c)
+	}
+}
+
+func sendResponsesStreamData(c *gin.Context, streamResponse dto.ResponsesStreamResponse, data string) {
+	if data == "" {
+		return
+	}
+	_ = helper.ResponseChunkData(c, streamResponse, data)
+}
@@ -0,0 +1,98 @@
+package openai
+
+import (
+	"bytes"
+	"io"
+	"mime/multipart"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/QuantumNous/new-api/common"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	relayconstant "github.com/QuantumNous/new-api/relay/constant"
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/gin-gonic/gin"
+	"github.com/stretchr/testify/require"
+)
+
+// TestConvertImageEditRequestMultipart verifies that ConvertImageRequest
+// re-serializes multipart image edit requests with all fields (including
+// stream) and the file intact, both when the form was already parsed and when
+// it must be re-parsed from the reusable body.
+func TestConvertImageEditRequestMultipart(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	newMultipartContext := func(t *testing.T, prompt string) *gin.Context {
+		var body bytes.Buffer
+		writer := multipart.NewWriter(&body)
+		require.NoError(t, writer.WriteField("model", "gpt-image-1"))
+		require.NoError(t, writer.WriteField("prompt", prompt))
+		require.NoError(t, writer.WriteField("stream", "true"))
+		require.NoError(t, writer.WriteField("partial_images", "3"))
+		part, err := writer.CreateFormFile("image", "input.png")
+		require.NoError(t, err)
+		_, err = part.Write([]byte("fake image"))
+		require.NoError(t, err)
+		require.NoError(t, writer.Close())
+
+		c, _ := gin.CreateTestContext(httptest.NewRecorder())
+		c.Request = httptest.NewRequest(http.MethodPost, "/v1/images/edits", &body)
+		c.Request.Header.Set("Content-Type", writer.FormDataContentType())
+		return c
+	}
+
+	convertAndReplay := func(t *testing.T, c *gin.Context, prompt string) {
+		info := &relaycommon.RelayInfo{
+			RelayMode: relayconstant.RelayModeImagesEdits,
+		}
+		request := dto.ImageRequest{
+			Model:  "gpt-image-1",
+			Prompt: prompt,
+			Stream: common.GetPointer(true),
+		}
+
+		converted, err := (&Adaptor{}).ConvertImageRequest(c, info, request)
+		require.NoError(t, err)
+		convertedBody, ok := converted.(*bytes.Buffer)
+		require.True(t, ok)
+
+		replayedRequest := httptest.NewRequest(http.MethodPost, "/v1/images/edits", bytes.NewReader(convertedBody.Bytes()))
+		replayedRequest.Header.Set("Content-Type", c.Request.Header.Get("Content-Type"))
+		require.NoError(t, replayedRequest.ParseMultipartForm(32<<20))
+
+		require.Equal(t, "gpt-image-1", replayedRequest.PostForm.Get("model"))
+		require.Equal(t, prompt, replayedRequest.PostForm.Get("prompt"))
+		require.Equal(t, "true", replayedRequest.PostForm.Get("stream"))
+		require.Equal(t, "3", replayedRequest.PostForm.Get("partial_images"))
+		require.Len(t, replayedRequest.MultipartForm.File["image"], 1)
+
+		file, err := replayedRequest.MultipartForm.File["image"][0].Open()
+		require.NoError(t, err)
+		defer file.Close()
+		fileBytes, err := io.ReadAll(file)
+		require.NoError(t, err)
+		require.Equal(t, []byte("fake image"), fileBytes)
+	}
+
+	t.Run("with pre-parsed form", func(t *testing.T) {
+		prompt := "edit this image"
+		c := newMultipartContext(t, prompt)
+		require.NoError(t, c.Request.ParseMultipartForm(32<<20))
+
+		convertAndReplay(t, c, prompt)
+	})
+
+	t.Run("re-parses reusable body when form is missing", func(t *testing.T) {
+		prompt := "edit without pre-parsed form"
+		c := newMultipartContext(t, prompt)
+
+		storage, err := common.GetBodyStorage(c)
+		require.NoError(t, err)
+		c.Request.Body = io.NopCloser(storage)
+		c.Request.MultipartForm = nil
+		c.Request.PostForm = nil
+
+		convertAndReplay(t, c, prompt)
+	})
+}
@@ -0,0 +1,449 @@
+package openai
+
+import (
+	"context"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"sync"
+	"testing"
+
+	"github.com/QuantumNous/new-api/constant"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	relayconstant "github.com/QuantumNous/new-api/relay/constant"
+	"github.com/gin-gonic/gin"
+	"github.com/stretchr/testify/require"
+)
+
+func newImageTestContext(t *testing.T, body, contentType string, isStream bool) (*gin.Context, *httptest.ResponseRecorder, *http.Response, *relaycommon.RelayInfo) {
+	t.Helper()
+
+	recorder := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(recorder)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/images/generations", nil)
+
+	resp := &http.Response{
+		StatusCode: http.StatusOK,
+		Body:       io.NopCloser(strings.NewReader(body)),
+		Header:     http.Header{"Content-Type": []string{contentType}},
+	}
+	info := &relaycommon.RelayInfo{
+		ChannelMeta: &relaycommon.ChannelMeta{},
+		IsStream:    isStream,
+	}
+	return c, recorder, resp, info
+}
+
+func TestOpenaiImageDoResponseUsesInfoIsStream(t *testing.T) {
+	oldMode := gin.Mode()
+	gin.SetMode(gin.TestMode)
+	t.Cleanup(func() { gin.SetMode(oldMode) })
+
+	body := `{"created":1710000000,"data":[{"b64_json":"image"}]}`
+
+	t.Run("non-stream response stays JSON", func(t *testing.T) {
+		c, recorder, resp, info := newImageTestContext(t, body, "application/json", false)
+		info.RelayMode = relayconstant.RelayModeImagesGenerations
+
+		usage, err := (&Adaptor{}).DoResponse(c, resp, info)
+
+		require.Nil(t, err)
+		require.NotNil(t, usage)
+		require.Equal(t, body, recorder.Body.String())
+	})
+
+	t.Run("stream response converts JSON to SSE", func(t *testing.T) {
+		c, recorder, resp, info := newImageTestContext(t, body, "application/json", true)
+		info.RelayMode = relayconstant.RelayModeImagesGenerations
+
+		usage, err := (&Adaptor{}).DoResponse(c, resp, info)
+
+		require.Nil(t, err)
+		require.NotNil(t, usage)
+		require.Contains(t, recorder.Body.String(), `event: image_generation.completed`)
+		require.Contains(t, recorder.Body.String(), `data: [DONE]`)
+	})
+}
+
+// TestOpenaiImageStreamHandlerForwardsSSEAndUsage covers the core SSE path:
+// chunks are forwarded with rebuilt event lines, usage is extracted and
+// normalized (input_tokens -> prompt_tokens with details), and [DONE] is
+// re-emitted to the client.
+func TestOpenaiImageStreamHandlerForwardsSSEAndUsage(t *testing.T) {
+	oldMode := gin.Mode()
+	gin.SetMode(gin.TestMode)
+	t.Cleanup(func() { gin.SetMode(oldMode) })
+
+	oldTimeout := constant.StreamingTimeout
+	constant.StreamingTimeout = 30
+	t.Cleanup(func() { constant.StreamingTimeout = oldTimeout })
+
+	body := strings.Join([]string{
+		`event: image_generation.partial_image`,
+		`data: {"type":"image_generation.partial_image","b64_json":"partial"}`,
+		``,
+		`data: {"usage":{"input_tokens":3,"output_tokens":4,"total_tokens":7,"input_tokens_details":{"image_tokens":2,"text_tokens":1}}}`,
+		``,
+		`data: [DONE]`,
+		``,
+	}, "\n")
+
+	c, recorder, resp, info := newImageTestContext(t, body, "text/event-stream", true)
+	info.PriceData.UsePrice = true
+	info.PriceData.AddOtherRatio("n", 3)
+
+	usage, err := OpenaiImageStreamHandler(c, info, resp)
+	require.Nil(t, err)
+	require.Equal(t, 3, usage.PromptTokens)
+	require.Equal(t, 4, usage.CompletionTokens)
+	require.Equal(t, 7, usage.TotalTokens)
+	require.Equal(t, 2, usage.PromptTokensDetails.ImageTokens)
+	require.Equal(t, 1, usage.PromptTokensDetails.TextTokens)
+	require.Contains(t, recorder.Body.String(), `event: image_generation.partial_image`)
+	require.Contains(t, recorder.Body.String(), `data: {"type":"image_generation.partial_image","b64_json":"partial"}`)
+	require.Contains(t, recorder.Body.String(), `data: {"usage":{"input_tokens":3,"output_tokens":4,"total_tokens":7,"input_tokens_details":{"image_tokens":2,"text_tokens":1}}}`)
+	require.Contains(t, recorder.Body.String(), `data: [DONE]`)
+	require.Equal(t, "text/event-stream", recorder.Header().Get("Content-Type"))
+	require.Equal(t, 3.0, info.PriceData.OtherRatios()["n"], "streams without completed events keep the requested count")
+}
+
+func TestOpenaiImageStreamHandlerUsesCompletedEventCount(t *testing.T) {
+	oldMode := gin.Mode()
+	gin.SetMode(gin.TestMode)
+	t.Cleanup(func() { gin.SetMode(oldMode) })
+
+	oldTimeout := constant.StreamingTimeout
+	constant.StreamingTimeout = 30
+	t.Cleanup(func() { constant.StreamingTimeout = oldTimeout })
+
+	body := strings.Join([]string{
+		`data: {"type":"image_generation.partial_image","partial_image_index":0,"b64_json":"partial"}`,
+		``,
+		`data: {"type":"image_generation.completed","b64_json":"first"}`,
+		``,
+		`data: {"type":"image_edit.completed","b64_json":"second","usage":{"input_tokens":3,"output_tokens":4,"total_tokens":7}}`,
+		``,
+		`data: [DONE]`,
+		``,
+	}, "\n")
+
+	c, _, resp, info := newImageTestContext(t, body, "text/event-stream", true)
+	info.PriceData.UsePrice = true
+	info.PriceData.AddOtherRatio("n", 3)
+
+	usage, err := OpenaiImageStreamHandler(c, info, resp)
+
+	require.Nil(t, err)
+	require.Equal(t, 7, usage.TotalTokens)
+	require.Equal(t, 2.0, info.PriceData.OtherRatios()["n"])
+}
+
+// blockingBody serves one SSE chunk, then blocks until Close (the scanner's
+// cleanup) and returns EOF — keeping the upstream "open" while the client-side
+// disconnect is simulated elsewhere.
+type blockingBody struct {
+	mu     sync.Mutex
+	sent   bool
+	chunk  []byte
+	closed chan struct{}
+}
+
+func (b *blockingBody) Read(p []byte) (int, error) {
+	b.mu.Lock()
+	if !b.sent {
+		b.sent = true
+		n := copy(p, b.chunk)
+		b.mu.Unlock()
+		return n, nil
+	}
+	b.mu.Unlock()
+	<-b.closed
+	return 0, io.EOF
+}
+
+func (b *blockingBody) Close() error {
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	select {
+	case <-b.closed:
+	default:
+		close(b.closed)
+	}
+	return nil
+}
+
+// cancelAfterWriter cancels the request context right after the payload
+// containing needle has been written to the client, simulating a client that
+// disconnects after receiving that event. Cancelling from the write side (not
+// the upstream read side) makes the abort deterministic: the handler has
+// already processed and counted the event when the disconnect fires.
+type cancelAfterWriter struct {
+	gin.ResponseWriter
+	needle string
+	cancel context.CancelFunc
+	once   sync.Once
+}
+
+func (w *cancelAfterWriter) Write(p []byte) (int, error) {
+	n, err := w.ResponseWriter.Write(p)
+	if strings.Contains(string(p), w.needle) {
+		w.once.Do(w.cancel)
+	}
+	return n, err
+}
+
+func (w *cancelAfterWriter) WriteString(s string) (int, error) {
+	n, err := io.WriteString(w.ResponseWriter, s)
+	if strings.Contains(s, w.needle) {
+		w.once.Do(w.cancel)
+	}
+	return n, err
+}
+
+func newDisconnectingImageStream(t *testing.T, sseBody, disconnectAfter string) (*gin.Context, *httptest.ResponseRecorder, *http.Response, *relaycommon.RelayInfo) {
+	t.Helper()
+	c, recorder, resp, info := newImageTestContext(t, "", "text/event-stream", true)
+	ctx, cancel := context.WithCancel(c.Request.Context())
+	t.Cleanup(cancel)
+	c.Request = c.Request.WithContext(ctx)
+	c.Writer = &cancelAfterWriter{ResponseWriter: c.Writer, needle: disconnectAfter, cancel: cancel}
+	resp.Body = &blockingBody{
+		chunk:  []byte(sseBody),
+		closed: make(chan struct{}),
+	}
+	return c, recorder, resp, info
+}
+
+// TestOpenaiImageStreamHandlerClientDisconnectKeepsRequestedCount guards the
+// billing invariant: completed-event counting must not lower the charge when
+// the client aborts the stream. Upstream already generated (and charged for)
+// all requested images, so a disconnect after the first completed event keeps
+// the requested n instead of dropping it to 1.
+func TestOpenaiImageStreamHandlerClientDisconnectKeepsRequestedCount(t *testing.T) {
+	oldMode := gin.Mode()
+	gin.SetMode(gin.TestMode)
+	t.Cleanup(func() { gin.SetMode(oldMode) })
+
+	oldTimeout := constant.StreamingTimeout
+	constant.StreamingTimeout = 30
+	t.Cleanup(func() { constant.StreamingTimeout = oldTimeout })
+
+	body := "data: {\"type\":\"image_generation.completed\",\"b64_json\":\"first\"}\n\n"
+	c, recorder, resp, info := newDisconnectingImageStream(t, body, "first")
+	info.PriceData.UsePrice = true
+	info.PriceData.AddOtherRatio("n", 3)
+
+	usage, err := OpenaiImageStreamHandler(c, info, resp)
+
+	require.Nil(t, err)
+	require.NotNil(t, usage)
+	require.NotNil(t, info.StreamStatus)
+	// A client abort surfaces as client_gone (main-loop ctx watch) or
+	// handler_stop (failed client write); both must be treated as untrusted.
+	require.Contains(t,
+		[]relaycommon.StreamEndReason{relaycommon.StreamEndReasonClientGone, relaycommon.StreamEndReasonHandlerStop},
+		info.StreamStatus.EndReason)
+	require.Contains(t, recorder.Body.String(), `"b64_json":"first"`)
+	require.Equal(t, 3.0, info.PriceData.OtherRatios()["n"], "client abort must not reduce the billed image count")
+}
+
+// TestOpenaiImageStreamHandlerClientDisconnectRaisesCount covers the other
+// direction of the abort guard: when completed events already exceed the
+// recorded n, the higher actual count is billed even though the client aborted.
+func TestOpenaiImageStreamHandlerClientDisconnectRaisesCount(t *testing.T) {
+	oldMode := gin.Mode()
+	gin.SetMode(gin.TestMode)
+	t.Cleanup(func() { gin.SetMode(oldMode) })
+
+	oldTimeout := constant.StreamingTimeout
+	constant.StreamingTimeout = 30
+	t.Cleanup(func() { constant.StreamingTimeout = oldTimeout })
+
+	body := strings.Join([]string{
+		`data: {"type":"image_generation.completed","b64_json":"first"}`,
+		``,
+		`data: {"type":"image_generation.completed","b64_json":"second"}`,
+		``,
+		``,
+	}, "\n")
+	c, _, resp, info := newDisconnectingImageStream(t, body, "second")
+	info.PriceData.UsePrice = true
+	info.PriceData.AddOtherRatio("n", 1)
+
+	usage, err := OpenaiImageStreamHandler(c, info, resp)
+
+	require.Nil(t, err)
+	require.NotNil(t, usage)
+	require.NotNil(t, info.StreamStatus)
+	require.Contains(t,
+		[]relaycommon.StreamEndReason{relaycommon.StreamEndReasonClientGone, relaycommon.StreamEndReasonHandlerStop},
+		info.StreamStatus.EndReason)
+	require.Equal(t, 2.0, info.PriceData.OtherRatios()["n"], "completed events beyond the recorded n must raise the charge even on abort")
+}
+
+// TestOpenaiImageStreamHandlerWrapsJSONResponse covers the non-SSE fallback:
+// a JSON upstream response is wrapped into pseudo-SSE completed events.
+func TestOpenaiImageStreamHandlerWrapsJSONResponse(t *testing.T) {
+	oldMode := gin.Mode()
+	gin.SetMode(gin.TestMode)
+	t.Cleanup(func() { gin.SetMode(oldMode) })
+
+	body := `{"created":1710000000,"data":[{"b64_json":"first","revised_prompt":"draw a cat"},{"b64_json":"second"}],"usage":{"input_tokens":3,"output_tokens":4,"total_tokens":7,"input_tokens_details":{"image_tokens":2,"text_tokens":1}}}`
+
+	c, recorder, resp, info := newImageTestContext(t, body, "application/json", true)
+	info.PriceData.UsePrice = true
+	info.PriceData.AddOtherRatio("n", 3)
+
+	usage, err := OpenaiImageStreamHandler(c, info, resp)
+	require.Nil(t, err)
+	require.Equal(t, 3, usage.PromptTokens)
+	require.Equal(t, 4, usage.CompletionTokens)
+	require.Equal(t, 7, usage.TotalTokens)
+	require.Equal(t, 2, usage.PromptTokensDetails.ImageTokens)
+	require.Equal(t, 1, usage.PromptTokensDetails.TextTokens)
+	require.Equal(t, "text/event-stream", recorder.Header().Get("Content-Type"))
+	require.Empty(t, recorder.Header().Get("Content-Length"))
+	require.Contains(t, recorder.Body.String(), `event: image_generation.completed`)
+	require.Contains(t, recorder.Body.String(), `"type":"image_generation.completed"`)
+	require.Contains(t, recorder.Body.String(), `"b64_json":"first"`)
+	require.Contains(t, recorder.Body.String(), `"b64_json":"second"`)
+	require.Contains(t, recorder.Body.String(), `"revised_prompt":"draw a cat"`)
+	require.Contains(t, recorder.Body.String(), `data: [DONE]`)
+	require.Equal(t, 2, strings.Count(recorder.Body.String(), `event: image_generation.completed`))
+	require.Equal(t, 2.0, info.PriceData.OtherRatios()["n"])
+}
+
+func TestOpenaiImageHandlerUsesPositiveActualCountForFixedPrice(t *testing.T) {
+	oldMode := gin.Mode()
+	gin.SetMode(gin.TestMode)
+	t.Cleanup(func() { gin.SetMode(oldMode) })
+	longImage := strings.Repeat("a", 4096)
+
+	tests := []struct {
+		name      string
+		body      string
+		usePrice  bool
+		wantCount float64
+	}{
+		{
+			name:      "fixed price uses data length",
+			body:      `{"data":[{"b64_json":"` + longImage + `"},{"b64_json":"second"}]}`,
+			usePrice:  true,
+			wantCount: 2,
+		},
+		{
+			name:      "empty data keeps requested count",
+			body:      `{"data":[]}`,
+			usePrice:  true,
+			wantCount: 3,
+		},
+		{
+			name:      "ratio billing ignores data length",
+			body:      `{"data":[{"b64_json":"first"},{"b64_json":"second"}]}`,
+			usePrice:  false,
+			wantCount: 3,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			c, recorder, resp, info := newImageTestContext(t, tt.body, "application/json", false)
+			info.PriceData.UsePrice = tt.usePrice
+			info.PriceData.AddOtherRatio("n", 3)
+
+			_, err := OpenaiImageHandler(c, info, resp)
+
+			require.Nil(t, err)
+			require.Equal(t, tt.wantCount, info.PriceData.OtherRatios()["n"])
+			require.Equal(t, tt.body, recorder.Body.String())
+		})
+	}
+}
+
+// TestOpenaiImageHandlersReturnJSONError covers JSON error responses for both
+// entry points: the non-streaming handler and the stream handler's non-SSE
+// fallback. Neither must leak the error body to the client.
+func TestOpenaiImageHandlersReturnJSONError(t *testing.T) {
+	oldMode := gin.Mode()
+	gin.SetMode(gin.TestMode)
+	t.Cleanup(func() { gin.SetMode(oldMode) })
+
+	body := `{"error":{"message":"content moderation failed","type":"upstream_error","code":"content_moderation_failed","status":502}}`
+
+	t.Run("non-streaming handler", func(t *testing.T) {
+		c, recorder, resp, info := newImageTestContext(t, body, "application/json", false)
+
+		usage, err := OpenaiImageHandler(c, info, resp)
+		require.Nil(t, usage)
+		require.NotNil(t, err)
+		require.Equal(t, http.StatusOK, err.StatusCode)
+		oaiError := err.ToOpenAIError()
+		require.Equal(t, "content moderation failed", oaiError.Message)
+		require.Equal(t, "upstream_error", oaiError.Type)
+		require.Equal(t, "content_moderation_failed", oaiError.Code)
+		require.Empty(t, recorder.Body.String())
+	})
+
+	t.Run("stream handler JSON fallback", func(t *testing.T) {
+		c, recorder, resp, info := newImageTestContext(t, body, "application/json", true)
+
+		usage, err := OpenaiImageStreamHandler(c, info, resp)
+		require.Nil(t, usage)
+		require.NotNil(t, err)
+		require.Equal(t, http.StatusOK, err.StatusCode)
+		require.Equal(t, "content moderation failed", err.ToOpenAIError().Message)
+		require.Empty(t, recorder.Body.String())
+	})
+
+	t.Run("stream handler non-2xx stays JSON error", func(t *testing.T) {
+		c, recorder, resp, info := newImageTestContext(t, body, "application/json", true)
+		resp.StatusCode = http.StatusBadGateway
+
+		usage, err := OpenaiImageStreamHandler(c, info, resp)
+		require.Nil(t, usage)
+		require.NotNil(t, err)
+		require.Equal(t, http.StatusBadGateway, err.StatusCode)
+		require.Equal(t, "content moderation failed", err.ToOpenAIError().Message)
+		require.Empty(t, recorder.Body.String())
+		require.NotContains(t, recorder.Header().Get("Content-Type"), "text/event-stream")
+	})
+}
+
+// TestOpenaiImageStreamHandlerRecordsUpstreamErrorEvent verifies that an error
+// event inside the SSE stream is recorded as a soft error while the payload is
+// still forwarded to the client.
+func TestOpenaiImageStreamHandlerRecordsUpstreamErrorEvent(t *testing.T) {
+	oldMode := gin.Mode()
+	gin.SetMode(gin.TestMode)
+	t.Cleanup(func() { gin.SetMode(oldMode) })
+
+	oldTimeout := constant.StreamingTimeout
+	constant.StreamingTimeout = 30
+	t.Cleanup(func() { constant.StreamingTimeout = oldTimeout })
+
+	body := strings.Join([]string{
+		`event: image_generation.partial_image`,
+		`data: {"type":"image_generation.partial_image","b64_json":"partial"}`,
+		``,
+		`event: error`,
+		`data: {"type":"upstream_error","error":{"message":"stream error: stream ID 77; INTERNAL_ERROR; received from peer"}}`,
+		``,
+	}, "\n")
+
+	c, recorder, resp, info := newImageTestContext(t, body, "text/event-stream", true)
+
+	usage, err := OpenaiImageStreamHandler(c, info, resp)
+	require.Nil(t, err)
+	require.NotNil(t, usage)
+	require.NotNil(t, info.StreamStatus)
+	require.Equal(t, relaycommon.StreamEndReasonEOF, info.StreamStatus.EndReason)
+	require.True(t, info.StreamStatus.HasErrors())
+	require.Equal(t, 1, info.StreamStatus.TotalErrorCount())
+	require.Contains(t, info.StreamStatus.Errors[0].Message, "INTERNAL_ERROR")
+	// The scanner strips the upstream "event: error" line; the event name is
+	// rebuilt from the JSON "type" field (upstream_error). The error message
+	// is still forwarded in the data: payload (stream ID 77).
+	require.Contains(t, recorder.Body.String(), `event: upstream_error`)
+	require.Contains(t, recorder.Body.String(), `stream ID 77`)
+}
@@ -0,0 +1,337 @@
+package openai
+
+import (
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	"github.com/QuantumNous/new-api/logger"
+	"github.com/QuantumNous/new-api/relay/channel/openrouter"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/QuantumNous/new-api/relay/helper"
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/QuantumNous/new-api/relaykit/relayconvert"
+	"github.com/QuantumNous/new-api/relaykit/types"
+	"github.com/QuantumNous/new-api/service"
+
+	"github.com/gin-gonic/gin"
+)
+
+func sendStreamData(c *gin.Context, info *relaycommon.RelayInfo, data string, forceFormat bool, thinkToContent bool) error {
+	if data == "" {
+		return nil
+	}
+
+	if !forceFormat && !thinkToContent {
+		return helper.StringData(c, data)
+	}
+
+	var lastStreamResponse dto.ChatCompletionsStreamResponse
+	if err := common.UnmarshalJsonStr(data, &lastStreamResponse); err != nil {
+		return err
+	}
+
+	if !thinkToContent {
+		return helper.ObjectData(c, lastStreamResponse)
+	}
+
+	hasThinkingContent := false
+	hasContent := false
+	var thinkingContent strings.Builder
+	for _, choice := range lastStreamResponse.Choices {
+		if len(choice.Delta.GetReasoningContent()) > 0 {
+			hasThinkingContent = true
+			thinkingContent.WriteString(choice.Delta.GetReasoningContent())
+		}
+		if len(choice.Delta.GetContentString()) > 0 {
+			hasContent = true
+		}
+	}
+
+	// Handle think to content conversion
+	if info.ThinkingContentInfo.IsFirstThinkingContent {
+		if hasThinkingContent {
+			response := lastStreamResponse.Copy()
+			for i := range response.Choices {
+				// send `think` tag with thinking content
+				response.Choices[i].Delta.SetContentString("<think>\n" + thinkingContent.String())
+				response.Choices[i].Delta.ReasoningContent = nil
+				response.Choices[i].Delta.Reasoning = nil
+			}
+			info.ThinkingContentInfo.IsFirstThinkingContent = false
+			info.ThinkingContentInfo.HasSentThinkingContent = true
+			return helper.ObjectData(c, response)
+		}
+	}
+
+	if lastStreamResponse.Choices == nil || len(lastStreamResponse.Choices) == 0 {
+		return helper.ObjectData(c, lastStreamResponse)
+	}
+
+	// Process each choice
+	for i, choice := range lastStreamResponse.Choices {
+		// Handle transition from thinking to content
+		// only send `</think>` tag when previous thinking content has been sent
+		if hasContent && !info.ThinkingContentInfo.SendLastThinkingContent && info.ThinkingContentInfo.HasSentThinkingContent {
+			response := lastStreamResponse.Copy()
+			for j := range response.Choices {
+				response.Choices[j].Delta.SetContentString("\n</think>\n")
+				response.Choices[j].Delta.ReasoningContent = nil
+				response.Choices[j].Delta.Reasoning = nil
+			}
+			info.ThinkingContentInfo.SendLastThinkingContent = true
+			helper.ObjectData(c, response)
+		}
+
+		// Convert reasoning content to regular content if any
+		if len(choice.Delta.GetReasoningContent()) > 0 {
+			lastStreamResponse.Choices[i].Delta.SetContentString(choice.Delta.GetReasoningContent())
+			lastStreamResponse.Choices[i].Delta.ReasoningContent = nil
+			lastStreamResponse.Choices[i].Delta.Reasoning = nil
+		} else if !hasThinkingContent && !hasContent {
+			// flush thinking content
+			lastStreamResponse.Choices[i].Delta.ReasoningContent = nil
+			lastStreamResponse.Choices[i].Delta.Reasoning = nil
+		}
+	}
+
+	return helper.ObjectData(c, lastStreamResponse)
+}
+
+func OaiStreamHandler(c *gin.Context, info *relaycommon.RelayInfo, resp *http.Response) (*dto.Usage, *types.NewAPIError) {
+	if resp == nil || resp.Body == nil {
+		logger.LogError(c, "invalid response or response body")
+		return nil, types.NewOpenAIError(fmt.Errorf("invalid response"), types.ErrorCodeBadResponse, http.StatusInternalServerError)
+	}
+
+	defer service.CloseResponseBodyGracefully(resp)
+
+	model := info.UpstreamModelName
+	var responseId string
+	var createAt int64 = 0
+	var systemFingerprint string
+	var containStreamUsage bool
+	var responseTextBuilder strings.Builder
+	var toolCount int
+	var usage = &dto.Usage{}
+	var lastStreamData string
+	var secondLastStreamData string // 存储倒数第二个stream data，用于音频模型
+	seenStreamToolCalls := make(map[string]struct{})
+	var streamFunctionCallNames []string
+
+	// 检查是否为音频模型
+	isAudioModel := strings.Contains(strings.ToLower(model), "audio")
+
+	helper.StreamScannerHandler(c, resp, info, func(data string, sr *helper.StreamResult) {
+		if lastStreamData != "" {
+			if err := HandleStreamFormat(c, info, lastStreamData, info.ChannelSetting.ForceFormat, info.ChannelSetting.ThinkingToContent); err != nil {
+				common.SysLog("error handling stream format: " + err.Error())
+				sr.Error(err)
+			}
+		}
+		if len(data) > 0 {
+			// 对音频模型，保存倒数第二个stream data
+			if isAudioModel && lastStreamData != "" {
+				secondLastStreamData = lastStreamData
+			}
+
+			lastStreamData = data
+			collectStreamFunctionCallNames(data, seenStreamToolCalls, &streamFunctionCallNames)
+			if err := processTokenData(info.RelayMode, data, &responseTextBuilder, &toolCount); err != nil {
+				logger.LogError(c, "error processing stream token data: "+err.Error())
+				sr.Error(err)
+			}
+		}
+	})
+
+	// 对音频模型，从倒数第二个stream data中提取usage信息
+	if isAudioModel && secondLastStreamData != "" {
+		var streamResp struct {
+			Usage *dto.Usage `json:"usage"`
+		}
+		err := common.Unmarshal([]byte(secondLastStreamData), &streamResp)
+		if err == nil && streamResp.Usage != nil && service.ValidUsage(streamResp.Usage) {
+			usage = streamResp.Usage
+			containStreamUsage = true
+
+			if common.DebugEnabled {
+				logger.LogDebug(c, "Audio model usage extracted from second last SSE: PromptTokens=%d, CompletionTokens=%d, TotalTokens=%d, InputTokens=%d, OutputTokens=%d",
+					usage.PromptTokens, usage.CompletionTokens, usage.TotalTokens,
+					usage.InputTokens, usage.OutputTokens)
+			}
+		}
+	}
+
+	// 处理最后的响应
+	shouldSendLastResp := true
+	if err := handleLastResponse(lastStreamData, &responseId, &createAt, &systemFingerprint, &model, &usage,
+		&containStreamUsage, info, &shouldSendLastResp); err != nil {
+		logger.LogError(c, fmt.Sprintf("error handling last response: %s, lastStreamData: [%s]", err.Error(), lastStreamData))
+	}
+
+	if info.RelayFormat == types.RelayFormatOpenAI {
+		if shouldSendLastResp {
+			_ = sendStreamData(c, info, lastStreamData, info.ChannelSetting.ForceFormat, info.ChannelSetting.ThinkingToContent)
+		}
+	}
+
+	if !containStreamUsage {
+		usage = service.ResponseText2Usage(c, responseTextBuilder.String(), info.UpstreamModelName, info.GetEstimatePromptTokens())
+		usage.CompletionTokens += toolCount * 7
+	}
+
+	applyUsagePostProcessing(info, usage, common.StringToByteSlice(lastStreamData))
+
+	for _, name := range streamFunctionCallNames {
+		info.CountBillableToolCall(dto.BuildInCallFunctionCall, name)
+	}
+
+	HandleFinalResponse(c, info, lastStreamData, responseId, createAt, model, systemFingerprint, usage, containStreamUsage)
+
+	return usage, nil
+}
+
+func collectStreamFunctionCallNames(data string, seen map[string]struct{}, names *[]string) {
+	var streamResponse dto.ChatCompletionsStreamResponse
+	if err := common.UnmarshalJsonStr(data, &streamResponse); err != nil {
+		return
+	}
+	for _, choice := range streamResponse.Choices {
+		for i, tc := range choice.Delta.ToolCalls {
+			name := tc.Function.Name
+			if name == "" {
+				continue
+			}
+			toolIdx := i
+			if tc.Index != nil {
+				toolIdx = *tc.Index
+			}
+			key := fmt.Sprintf("%d-%d", choice.Index, toolIdx)
+			if _, ok := seen[key]; ok {
+				continue
+			}
+			seen[key] = struct{}{}
+			*names = append(*names, name)
+		}
+	}
+}
+
+func OpenaiHandler(c *gin.Context, info *relaycommon.RelayInfo, resp *http.Response) (*dto.Usage, *types.NewAPIError) {
+	defer service.CloseResponseBodyGracefully(resp)
+
+	var simpleResponse dto.OpenAITextResponse
+	responseBody, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeReadResponseBodyFailed, http.StatusInternalServerError)
+	}
+	logger.LogDebug(c, "upstream response body: %s", responseBody)
+	// Unmarshal to simpleResponse
+	if info.ChannelType == constant.ChannelTypeOpenRouter && info.ChannelOtherSettings.IsOpenRouterEnterprise() {
+		// 尝试解析为 openrouter enterprise
+		var enterpriseResponse openrouter.OpenRouterEnterpriseResponse
+		err = common.Unmarshal(responseBody, &enterpriseResponse)
+		if err != nil {
+			return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+		}
+		if enterpriseResponse.Success {
+			responseBody = enterpriseResponse.Data
+		} else {
+			logger.LogError(c, fmt.Sprintf("openrouter enterprise response success=false, data: %s", enterpriseResponse.Data))
+			return nil, types.NewOpenAIError(fmt.Errorf("openrouter response success=false"), types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+		}
+	}
+
+	err = common.Unmarshal(responseBody, &simpleResponse)
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+	}
+
+	if oaiError := simpleResponse.GetOpenAIError(); oaiError != nil && oaiError.Type != "" {
+		return nil, types.WithOpenAIError(*oaiError, resp.StatusCode)
+	}
+
+	for _, choice := range simpleResponse.Choices {
+		if choice.FinishReason == constant.FinishReasonContentFilter {
+			common.SetContextKey(c, constant.ContextKeyAdminRejectReason, "openai_finish_reason=content_filter")
+			break
+		}
+	}
+
+	for _, choice := range simpleResponse.Choices {
+		for _, tc := range choice.Message.ParseToolCalls() {
+			info.CountBillableToolCall(dto.BuildInCallFunctionCall, tc.Function.Name)
+		}
+	}
+
+	forceFormat := false
+	if info.ChannelSetting.ForceFormat {
+		forceFormat = true
+	}
+
+	usageModified := false
+	if simpleResponse.Usage.PromptTokens == 0 {
+		completionTokens := simpleResponse.Usage.CompletionTokens
+		if completionTokens == 0 {
+			for _, choice := range simpleResponse.Choices {
+				ctkm := service.CountTextToken(choice.Message.StringContent()+choice.Message.GetReasoningContent(), info.UpstreamModelName)
+				completionTokens += ctkm
+			}
+		}
+		simpleResponse.Usage = dto.Usage{
+			PromptTokens:     info.GetEstimatePromptTokens(),
+			CompletionTokens: completionTokens,
+			TotalTokens:      info.GetEstimatePromptTokens() + completionTokens,
+		}
+		usageModified = true
+	}
+
+	applyUsagePostProcessing(info, &simpleResponse.Usage, responseBody)
+
+	switch info.RelayFormat {
+	case types.RelayFormatOpenAI:
+		if usageModified {
+			var bodyMap map[string]interface{}
+			err = common.Unmarshal(responseBody, &bodyMap)
+			if err != nil {
+				return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+			}
+			bodyMap["usage"] = simpleResponse.Usage
+			responseBody, _ = common.Marshal(bodyMap)
+		}
+		if forceFormat {
+			responseBody, err = common.Marshal(simpleResponse)
+			if err != nil {
+				return nil, types.NewError(err, types.ErrorCodeBadResponseBody)
+			}
+		} else {
+			break
+		}
+	case types.RelayFormatClaude:
+		convertResult, err := relayconvert.ConvertResponse(c, info, types.RelayFormatClaude, &simpleResponse)
+		if err != nil {
+			return nil, types.NewError(err, types.ErrorCodeBadResponseBody)
+		}
+		claudeRespStr, err := common.Marshal(convertResult.Value)
+		if err != nil {
+			return nil, types.NewError(err, types.ErrorCodeBadResponseBody)
+		}
+		responseBody = claudeRespStr
+	case types.RelayFormatGemini:
+		convertResult, err := relayconvert.ConvertResponse(c, info, types.RelayFormatGemini, &simpleResponse)
+		if err != nil {
+			return nil, types.NewError(err, types.ErrorCodeBadResponseBody)
+		}
+		geminiRespStr, err := common.Marshal(convertResult.Value)
+		if err != nil {
+			return nil, types.NewError(err, types.ErrorCodeBadResponseBody)
+		}
+		responseBody = geminiRespStr
+	}
+
+	service.IOCopyBytesGracefully(c, resp, responseBody)
+
+	return &simpleResponse.Usage, nil
+}
@@ -0,0 +1,333 @@
+package openai
+
+import (
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/logger"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/QuantumNous/new-api/relay/helper"
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/QuantumNous/new-api/relaykit/types"
+	"github.com/QuantumNous/new-api/service"
+
+	"github.com/gin-gonic/gin"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+func updateOpenAIImageCount(info *relaycommon.RelayInfo, count int64) {
+	if info == nil || !info.PriceData.UsePrice || count <= 0 || count > int64(dto.MaxImageN) {
+		return
+	}
+	info.PriceData.AddOtherRatio("n", float64(count))
+}
+
+// OpenaiImageHandler handles non-streaming OpenAI image responses
+// (generations/edits), returning the parsed usage for billing.
+func OpenaiImageHandler(c *gin.Context, info *relaycommon.RelayInfo, resp *http.Response) (*dto.Usage, *types.NewAPIError) {
+	defer service.CloseResponseBodyGracefully(resp)
+
+	responseBody, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeReadResponseBodyFailed, http.StatusInternalServerError)
+	}
+
+	var usageResp dto.SimpleResponse
+	err = common.Unmarshal(responseBody, &usageResp)
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+	}
+
+	if oaiError := usageResp.GetOpenAIError(); oaiError != nil && oaiError.Type != "" {
+		return nil, types.WithOpenAIError(*oaiError, resp.StatusCode)
+	}
+
+	updateOpenAIImageCount(info, gjson.GetBytes(responseBody, "data.#").Int())
+
+	// 写入新的 response body
+	service.IOCopyBytesGracefully(c, resp, responseBody)
+
+	normalizeOpenAIUsage(&usageResp.Usage)
+	applyUsagePostProcessing(info, &usageResp.Usage, responseBody)
+	return &usageResp.Usage, nil
+}
+
+// normalizeOpenAIUsage maps the OpenAI Images usage shape (input_tokens /
+// output_tokens / input_tokens_details) onto the canonical prompt/completion
+// fields. It is used only on the OpenAI image relay paths (generations/edits,
+// streaming and non-streaming): the image API never returns prompt_tokens /
+// completion_tokens, so the overwrite (=) semantics here are equivalent to the
+// previous additive (+=) behavior while avoiding any future double-counting if
+// both field sets are ever populated. Do not reuse this on chat/embedding paths
+// without revisiting the overwrite semantics.
+func normalizeOpenAIUsage(usage *dto.Usage) {
+	if usage == nil {
+		return
+	}
+	if usage.InputTokens != 0 {
+		usage.PromptTokens = usage.InputTokens
+	}
+	if usage.OutputTokens != 0 {
+		usage.CompletionTokens = usage.OutputTokens
+	}
+	if usage.InputTokensDetails != nil {
+		usage.PromptTokensDetails.CachedTokens = usage.InputTokensDetails.CachedTokens
+		usage.PromptTokensDetails.CachedCreationTokens = usage.InputTokensDetails.CachedCreationTokens
+		usage.PromptTokensDetails.CacheWriteTokens = usage.InputTokensDetails.CacheWriteTokens
+		usage.PromptTokensDetails.ImageTokens = usage.InputTokensDetails.ImageTokens
+		usage.PromptTokensDetails.TextTokens = usage.InputTokensDetails.TextTokens
+		usage.PromptTokensDetails.AudioTokens = usage.InputTokensDetails.AudioTokens
+	}
+	if usage.TotalTokens == 0 {
+		usage.TotalTokens = usage.PromptTokens + usage.CompletionTokens
+	}
+}
+
+func OpenaiImageStreamHandler(c *gin.Context, info *relaycommon.RelayInfo, resp *http.Response) (*dto.Usage, *types.NewAPIError) {
+	if resp == nil || resp.Body == nil {
+		logger.LogError(c, "invalid image stream response")
+		return nil, types.NewOpenAIError(fmt.Errorf("invalid response"), types.ErrorCodeBadResponse, http.StatusInternalServerError)
+	}
+
+	contentType := strings.ToLower(resp.Header.Get("Content-Type"))
+	if resp.StatusCode < http.StatusOK || resp.StatusCode >= http.StatusMultipleChoices {
+		return OpenaiImageHandler(c, info, resp)
+	}
+	if !strings.Contains(contentType, "text/event-stream") {
+		return openaiImageJSONAsStreamHandler(c, info, resp)
+	}
+	// Reuse the shared streaming engine (helper.StreamScannerHandler) so the
+	// image streaming path gets the same ping keepalive, streaming-timeout
+	// watchdog, client-disconnect detection, panic recovery and goroutine
+	// cleanup as every other relay stream. The scanner delivers only the
+	// "data:" payload, so the SSE "event:" line is rebuilt from the JSON "type"
+	// field (real OpenAI image events keep event == type).
+	usage := &dto.Usage{}
+	var lastStreamData []byte
+	var completedImages int64
+
+	helper.StreamScannerHandler(c, resp, info, func(data string, sr *helper.StreamResult) {
+		raw := common.StringToByteSlice(data)
+		lastStreamData = raw
+		if isOpenAIImageStreamErrorEvent(raw) {
+			// Record the error as a soft error; the scanner drives the final
+			// EndReason. HasErrors() flags the failure for logging/handling.
+			sr.Error(fmt.Errorf("%s", extractOpenAIImageStreamErrorMessage(raw)))
+		}
+		var chunk struct {
+			Type  string    `json:"type"`
+			Usage dto.Usage `json:"usage"`
+		}
+		if err := common.Unmarshal(raw, &chunk); err == nil {
+			normalizeOpenAIUsage(&chunk.Usage)
+			if service.ValidUsage(&chunk.Usage) {
+				usage = &chunk.Usage
+			}
+			if chunk.Type == "image_generation.completed" || chunk.Type == "image_edit.completed" {
+				completedImages++
+			}
+		}
+		if err := writeOpenaiImageStreamChunk(c, raw); err != nil {
+			sr.Stop(err)
+		}
+	})
+
+	// StreamScannerHandler consumes the upstream [DONE]; re-emit it so the
+	// client still receives a terminal data: [DONE].
+	if info.StreamStatus != nil && info.StreamStatus.EndReason == relaycommon.StreamEndReasonDone {
+		helper.Done(c)
+	}
+
+	applyUsagePostProcessing(info, usage, lastStreamData)
+	// Only trust completedImages when upstream finished the stream (done/eof).
+	// On client-side aborts (client_gone, or handler_stop from a failed client
+	// write) the counter undercounts what upstream actually generated and
+	// charged, so keep the requested n — otherwise a client could pay for one
+	// image by disconnecting right after the first completed event. The abort
+	// guard only blocks lowering the charge: if completed events already
+	// exceed the recorded n, bill the higher actual count regardless.
+	if info.StreamStatus != nil {
+		upstreamFinished := info.StreamStatus.EndReason == relaycommon.StreamEndReasonDone ||
+			info.StreamStatus.EndReason == relaycommon.StreamEndReasonEOF
+		requestedN := 1.0
+		if n, ok := info.PriceData.OtherRatios()["n"]; ok {
+			requestedN = n
+		}
+		if upstreamFinished || float64(completedImages) > requestedN {
+			updateOpenAIImageCount(info, completedImages)
+		}
+	}
+	return usage, nil
+}
+
+// writeOpenaiImageStreamChunk rebuilds the SSE frame for an image stream chunk:
+// it emits an "event:" line derived from the JSON "type" field (when present)
+// followed by the verbatim "data:" payload, mirroring helper.ResponseChunkData.
+func writeOpenaiImageStreamChunk(c *gin.Context, data []byte) error {
+	var payload struct {
+		Type string `json:"type"`
+	}
+	_ = common.Unmarshal(data, &payload)
+	if eventName := strings.TrimSpace(payload.Type); eventName != "" {
+		return helper.ResponseChunkData(c, dto.ResponsesStreamResponse{Type: eventName}, string(data))
+	}
+	return helper.StringData(c, string(data))
+}
+
+// isOpenAIImageStreamErrorEvent detects upstream error chunks by JSON content
+// only ("type" of error/upstream_error, or a non-empty "error" field). The SSE
+// "event:" line is not available here: StreamScannerHandler delivers only the
+// "data:" payload. A payload carrying just a "message" key is deliberately NOT
+// treated as an error to avoid false positives.
+func isOpenAIImageStreamErrorEvent(data []byte) bool {
+	if !json.Valid(data) {
+		return false
+	}
+	var payload struct {
+		Type  string          `json:"type"`
+		Error json.RawMessage `json:"error"`
+	}
+	if err := common.Unmarshal(data, &payload); err != nil {
+		return false
+	}
+	payloadType := strings.ToLower(strings.TrimSpace(payload.Type))
+	return payloadType == "error" || payloadType == "upstream_error" || len(payload.Error) > 0
+}
+
+func extractOpenAIImageStreamErrorMessage(data []byte) string {
+	if len(data) == 0 || !json.Valid(data) {
+		return "upstream image stream returned error event"
+	}
+	var payload struct {
+		Message string          `json:"message"`
+		Error   json.RawMessage `json:"error"`
+	}
+	if err := common.Unmarshal(data, &payload); err != nil {
+		return "upstream image stream returned error event"
+	}
+	if msg := strings.TrimSpace(payload.Message); msg != "" {
+		return msg
+	}
+	if len(payload.Error) > 0 {
+		var nested struct {
+			Message string `json:"message"`
+		}
+		if err := common.Unmarshal(payload.Error, &nested); err == nil {
+			if msg := strings.TrimSpace(nested.Message); msg != "" {
+				return msg
+			}
+		}
+		if msg := strings.TrimSpace(common.JsonRawMessageToString(payload.Error)); msg != "" {
+			return msg
+		}
+	}
+	return "upstream image stream returned error event"
+}
+
+func openaiImageJSONAsStreamHandler(c *gin.Context, info *relaycommon.RelayInfo, resp *http.Response) (*dto.Usage, *types.NewAPIError) {
+	defer service.CloseResponseBodyGracefully(resp)
+
+	responseBody, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeReadResponseBodyFailed, http.StatusInternalServerError)
+	}
+
+	// Only decode usage/error. Do not Unmarshal data[] into dto.ImageResponse —
+	// b64_json values are large and would be copied into Go strings then
+	// re-marshaled for each SSE event.
+	var usageResp dto.SimpleResponse
+	if err := common.Unmarshal(responseBody, &usageResp); err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+	}
+	if oaiError := usageResp.GetOpenAIError(); oaiError != nil && oaiError.Type != "" {
+		return nil, types.WithOpenAIError(*oaiError, resp.StatusCode)
+	}
+	normalizeOpenAIUsage(&usageResp.Usage)
+	applyUsagePostProcessing(info, &usageResp.Usage, responseBody)
+
+	imageCount := gjson.GetBytes(responseBody, "data.#").Int()
+	updateOpenAIImageCount(info, imageCount)
+
+	helper.SetEventStreamHeaders(c)
+	c.Status(http.StatusOK)
+
+	created := gjson.GetBytes(responseBody, "created").Int()
+	if created == 0 {
+		created = time.Now().Unix()
+	}
+	if info != nil {
+		info.SetFirstResponseTime()
+	}
+
+	validUsage := service.ValidUsage(&usageResp.Usage)
+	var usageJSON []byte
+	if validUsage {
+		usageJSON, err = common.Marshal(usageResp.Usage)
+		if err != nil {
+			return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+		}
+	}
+
+	for i := int64(0); i < imageCount; i++ {
+		image := gjson.GetBytes(responseBody, "data."+strconv.FormatInt(i, 10))
+		payload := []byte(`{"type":"image_generation.completed"}`)
+		payload, err = sjson.SetBytes(payload, "created_at", created)
+		if err != nil {
+			return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+		}
+		if validUsage {
+			payload, err = sjson.SetRawBytes(payload, "usage", usageJSON)
+			if err != nil {
+				return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+			}
+		}
+		// b64_json goes last: every sjson.Set* reallocates the whole payload,
+		// so inserting the large blob after all small fields avoids re-copying
+		// multi-MB buffers.
+		for _, field := range []string{"url", "revised_prompt", "b64_json"} {
+			value := image.Get(field)
+			if value.Type != gjson.String || value.Raw == `""` {
+				continue
+			}
+			raw := []byte(value.Raw)
+			if value.Index > 0 {
+				raw = responseBody[value.Index : value.Index+len(value.Raw)]
+			}
+			payload, err = sjson.SetRawBytes(payload, field, raw)
+			if err != nil {
+				return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+			}
+		}
+		if writeErr := helper.ResponseChunkData(c, dto.ResponsesStreamResponse{Type: "image_generation.completed"}, string(payload)); writeErr != nil {
+			if info != nil && info.StreamStatus != nil {
+				info.StreamStatus.SetEndReason(relaycommon.StreamEndReasonClientGone, writeErr)
+			}
+			return &usageResp.Usage, nil
+		}
+	}
+	if err := writeOpenaiImageStreamDone(c); err != nil {
+		if info != nil && info.StreamStatus != nil {
+			info.StreamStatus.SetEndReason(relaycommon.StreamEndReasonClientGone, err)
+		}
+		return &usageResp.Usage, nil
+	}
+	if info != nil {
+		info.ReceivedResponseCount += int(imageCount)
+		if info.StreamStatus == nil {
+			info.StreamStatus = relaycommon.NewStreamStatus()
+		}
+		info.StreamStatus.SetEndReason(relaycommon.StreamEndReasonDone, nil)
+	}
+	return &usageResp.Usage, nil
+}
+
+func writeOpenaiImageStreamDone(c *gin.Context) error {
+	return helper.StringData(c, "[DONE]")
+}
@@ -0,0 +1,242 @@
+package openai
+
+import (
+	"fmt"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/logger"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/QuantumNous/new-api/relay/helper"
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/QuantumNous/new-api/relaykit/types"
+	"github.com/QuantumNous/new-api/service"
+
+	"github.com/bytedance/gopkg/util/gopool"
+	"github.com/gin-gonic/gin"
+	"github.com/gorilla/websocket"
+)
+
+func OpenaiRealtimeHandler(c *gin.Context, info *relaycommon.RelayInfo) (*types.NewAPIError, *dto.RealtimeUsage) {
+	if info == nil || info.ClientWs == nil || info.TargetWs == nil {
+		return types.NewError(fmt.Errorf("invalid websocket connection"), types.ErrorCodeBadResponse), nil
+	}
+
+	info.IsStream = true
+	clientConn := info.ClientWs
+	targetConn := info.TargetWs
+
+	clientClosed := make(chan struct{})
+	targetClosed := make(chan struct{})
+	sendChan := make(chan []byte, 100)
+	receiveChan := make(chan []byte, 100)
+	errChan := make(chan error, 2)
+
+	usage := &dto.RealtimeUsage{}
+	localUsage := &dto.RealtimeUsage{}
+	sumUsage := &dto.RealtimeUsage{}
+
+	gopool.Go(func() {
+		defer func() {
+			if r := recover(); r != nil {
+				errChan <- fmt.Errorf("panic in client reader: %v", r)
+			}
+		}()
+		for {
+			select {
+			case <-c.Done():
+				return
+			default:
+				_, message, err := clientConn.ReadMessage()
+				if err != nil {
+					if !websocket.IsCloseError(err, websocket.CloseNormalClosure, websocket.CloseGoingAway) {
+						errChan <- fmt.Errorf("error reading from client: %v", err)
+					}
+					close(clientClosed)
+					return
+				}
+
+				realtimeEvent := &dto.RealtimeEvent{}
+				err = common.Unmarshal(message, realtimeEvent)
+				if err != nil {
+					errChan <- fmt.Errorf("error unmarshalling message: %v", err)
+					return
+				}
+
+				if realtimeEvent.Type == dto.RealtimeEventTypeSessionUpdate {
+					if realtimeEvent.Session != nil {
+						if realtimeEvent.Session.Tools != nil {
+							info.RealtimeTools = realtimeEvent.Session.Tools
+						}
+					}
+				}
+
+				textToken, audioToken, err := service.CountTokenRealtime(info, *realtimeEvent, info.UpstreamModelName)
+				if err != nil {
+					errChan <- fmt.Errorf("error counting text token: %v", err)
+					return
+				}
+				logger.LogInfo(c, fmt.Sprintf("type: %s, textToken: %d, audioToken: %d", realtimeEvent.Type, textToken, audioToken))
+				localUsage.TotalTokens += textToken + audioToken
+				localUsage.InputTokens += textToken + audioToken
+				localUsage.InputTokenDetails.TextTokens += textToken
+				localUsage.InputTokenDetails.AudioTokens += audioToken
+
+				err = helper.WssString(c, targetConn, string(message))
+				if err != nil {
+					errChan <- fmt.Errorf("error writing to target: %v", err)
+					return
+				}
+
+				select {
+				case sendChan <- message:
+				default:
+				}
+			}
+		}
+	})
+
+	gopool.Go(func() {
+		defer func() {
+			if r := recover(); r != nil {
+				errChan <- fmt.Errorf("panic in target reader: %v", r)
+			}
+		}()
+		for {
+			select {
+			case <-c.Done():
+				return
+			default:
+				_, message, err := targetConn.ReadMessage()
+				if err != nil {
+					if !websocket.IsCloseError(err, websocket.CloseNormalClosure, websocket.CloseGoingAway) {
+						errChan <- fmt.Errorf("error reading from target: %v", err)
+					}
+					close(targetClosed)
+					return
+				}
+				info.SetFirstResponseTime()
+				realtimeEvent := &dto.RealtimeEvent{}
+				err = common.Unmarshal(message, realtimeEvent)
+				if err != nil {
+					errChan <- fmt.Errorf("error unmarshalling message: %v", err)
+					return
+				}
+
+				if realtimeEvent.Type == dto.RealtimeEventTypeResponseDone {
+					realtimeUsage := realtimeEvent.Response.Usage
+					if realtimeUsage != nil {
+						usage.TotalTokens += realtimeUsage.TotalTokens
+						usage.InputTokens += realtimeUsage.InputTokens
+						usage.OutputTokens += realtimeUsage.OutputTokens
+						usage.InputTokenDetails.AudioTokens += realtimeUsage.InputTokenDetails.AudioTokens
+						usage.InputTokenDetails.CachedTokens += realtimeUsage.InputTokenDetails.CachedTokens
+						usage.InputTokenDetails.TextTokens += realtimeUsage.InputTokenDetails.TextTokens
+						usage.OutputTokenDetails.AudioTokens += realtimeUsage.OutputTokenDetails.AudioTokens
+						usage.OutputTokenDetails.TextTokens += realtimeUsage.OutputTokenDetails.TextTokens
+						err := preConsumeUsage(c, info, usage, sumUsage)
+						if err != nil {
+							errChan <- fmt.Errorf("error consume usage: %v", err)
+							return
+						}
+						// 本次计费完成，清除
+						usage = &dto.RealtimeUsage{}
+
+						localUsage = &dto.RealtimeUsage{}
+					} else {
+						textToken, audioToken, err := service.CountTokenRealtime(info, *realtimeEvent, info.UpstreamModelName)
+						if err != nil {
+							errChan <- fmt.Errorf("error counting text token: %v", err)
+							return
+						}
+						logger.LogInfo(c, fmt.Sprintf("type: %s, textToken: %d, audioToken: %d", realtimeEvent.Type, textToken, audioToken))
+						localUsage.TotalTokens += textToken + audioToken
+						info.IsFirstRequest = false
+						localUsage.InputTokens += textToken + audioToken
+						localUsage.InputTokenDetails.TextTokens += textToken
+						localUsage.InputTokenDetails.AudioTokens += audioToken
+						err = preConsumeUsage(c, info, localUsage, sumUsage)
+						if err != nil {
+							errChan <- fmt.Errorf("error consume usage: %v", err)
+							return
+						}
+						// 本次计费完成，清除
+						localUsage = &dto.RealtimeUsage{}
+						// print now usage
+					}
+					logger.LogInfo(c, fmt.Sprintf("realtime streaming sumUsage: %v", sumUsage))
+					logger.LogInfo(c, fmt.Sprintf("realtime streaming localUsage: %v", localUsage))
+					logger.LogInfo(c, fmt.Sprintf("realtime streaming localUsage: %v", localUsage))
+
+				} else if realtimeEvent.Type == dto.RealtimeEventTypeSessionUpdated || realtimeEvent.Type == dto.RealtimeEventTypeSessionCreated {
+					realtimeSession := realtimeEvent.Session
+					if realtimeSession != nil {
+						// update audio format
+						info.InputAudioFormat = common.GetStringIfEmpty(realtimeSession.InputAudioFormat, info.InputAudioFormat)
+						info.OutputAudioFormat = common.GetStringIfEmpty(realtimeSession.OutputAudioFormat, info.OutputAudioFormat)
+					}
+				} else {
+					textToken, audioToken, err := service.CountTokenRealtime(info, *realtimeEvent, info.UpstreamModelName)
+					if err != nil {
+						errChan <- fmt.Errorf("error counting text token: %v", err)
+						return
+					}
+					logger.LogInfo(c, fmt.Sprintf("type: %s, textToken: %d, audioToken: %d", realtimeEvent.Type, textToken, audioToken))
+					localUsage.TotalTokens += textToken + audioToken
+					localUsage.OutputTokens += textToken + audioToken
+					localUsage.OutputTokenDetails.TextTokens += textToken
+					localUsage.OutputTokenDetails.AudioTokens += audioToken
+				}
+
+				err = helper.WssString(c, clientConn, string(message))
+				if err != nil {
+					errChan <- fmt.Errorf("error writing to client: %v", err)
+					return
+				}
+
+				select {
+				case receiveChan <- message:
+				default:
+				}
+			}
+		}
+	})
+
+	select {
+	case <-clientClosed:
+	case <-targetClosed:
+	case err := <-errChan:
+		//return service.OpenAIErrorWrapper(err, "realtime_error", http.StatusInternalServerError), nil
+		logger.LogError(c, "realtime error: "+err.Error())
+	case <-c.Done():
+	}
+
+	if usage.TotalTokens != 0 {
+		_ = preConsumeUsage(c, info, usage, sumUsage)
+	}
+
+	if localUsage.TotalTokens != 0 {
+		_ = preConsumeUsage(c, info, localUsage, sumUsage)
+	}
+
+	// check usage total tokens, if 0, use local usage
+
+	return nil, sumUsage
+}
+
+func preConsumeUsage(ctx *gin.Context, info *relaycommon.RelayInfo, usage *dto.RealtimeUsage, totalUsage *dto.RealtimeUsage) error {
+	if usage == nil || totalUsage == nil {
+		return fmt.Errorf("invalid usage pointer")
+	}
+
+	totalUsage.TotalTokens += usage.TotalTokens
+	totalUsage.InputTokens += usage.InputTokens
+	totalUsage.OutputTokens += usage.OutputTokens
+	totalUsage.InputTokenDetails.CachedTokens += usage.InputTokenDetails.CachedTokens
+	totalUsage.InputTokenDetails.TextTokens += usage.InputTokenDetails.TextTokens
+	totalUsage.InputTokenDetails.AudioTokens += usage.InputTokenDetails.AudioTokens
+	totalUsage.OutputTokenDetails.TextTokens += usage.OutputTokenDetails.TextTokens
+	totalUsage.OutputTokenDetails.AudioTokens += usage.OutputTokenDetails.AudioTokens
+	// clear usage
+	err := service.PreWssConsumeQuota(ctx, info, usage)
+	return err
+}
@@ -0,0 +1,178 @@
+package openai
+
+import (
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/logger"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/QuantumNous/new-api/relay/helper"
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/QuantumNous/new-api/relaykit/types"
+	"github.com/QuantumNous/new-api/service"
+
+	"github.com/gin-gonic/gin"
+)
+
+func OaiResponsesHandler(c *gin.Context, info *relaycommon.RelayInfo, resp *http.Response) (*dto.Usage, *types.NewAPIError) {
+	defer service.CloseResponseBodyGracefully(resp)
+
+	// read response body
+	var responsesResponse dto.OpenAIResponsesResponse
+	responseBody, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeReadResponseBodyFailed, http.StatusInternalServerError)
+	}
+	err = common.Unmarshal(responseBody, &responsesResponse)
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+	}
+	if oaiError := responsesResponse.GetOpenAIError(); oaiError != nil && oaiError.Type != "" {
+		return nil, types.WithOpenAIError(*oaiError, resp.StatusCode)
+	}
+
+	// 写入新的 response body
+	service.IOCopyBytesGracefully(c, resp, responseBody)
+
+	// compute usage
+	usage := dto.Usage{}
+	if responsesResponse.Usage != nil {
+		usage.PromptTokens = responsesResponse.Usage.InputTokens
+		usage.CompletionTokens = responsesResponse.Usage.OutputTokens
+		usage.TotalTokens = responsesResponse.Usage.TotalTokens
+		if responsesResponse.Usage.InputTokensDetails != nil {
+			usage.PromptTokensDetails.CachedTokens = responsesResponse.Usage.InputTokensDetails.CachedTokens
+			usage.PromptTokensDetails.CacheWriteTokens = responsesResponse.Usage.InputTokensDetails.CacheWriteTokens
+		}
+	}
+	// Count actual tool invocations from Output (not tool declarations).
+	for _, output := range responsesResponse.Output {
+		switch output.Type {
+		case dto.BuildInCallWebSearchCall:
+			info.CountBillableToolCall(dto.BuildInCallWebSearchCall, "")
+		case dto.BuildInCallFileSearchCall:
+			info.CountBillableToolCall(dto.BuildInCallFileSearchCall, "")
+		case dto.BuildInCallFunctionCall:
+			info.CountBillableToolCall(dto.BuildInCallFunctionCall, output.Name)
+		}
+	}
+
+	imageCounter := &relaycommon.ImageGenerationCallCounter{}
+	if !relaycommon.IsNonBillableResponsesStatus(responsesResponse.Status) {
+		for i := range responsesResponse.Output {
+			idx := i
+			imageCounter.Observe(&responsesResponse.Output[i], &idx)
+		}
+	}
+	imageCounter.Commit(info)
+
+	return &usage, nil
+}
+
+func OaiResponsesStreamHandler(c *gin.Context, info *relaycommon.RelayInfo, resp *http.Response) (*dto.Usage, *types.NewAPIError) {
+	if resp == nil || resp.Body == nil {
+		logger.LogError(c, "invalid response or response body")
+		return nil, types.NewError(fmt.Errorf("invalid response"), types.ErrorCodeBadResponse)
+	}
+
+	defer service.CloseResponseBodyGracefully(resp)
+
+	var usage = &dto.Usage{}
+	var responseTextBuilder strings.Builder
+	imageCounter := &relaycommon.ImageGenerationCallCounter{}
+	imageCommitted := false
+
+	helper.StreamScannerHandler(c, resp, info, func(data string, sr *helper.StreamResult) {
+
+		// 检查当前数据是否包含 completed 状态和 usage 信息
+		var streamResponse dto.ResponsesStreamResponse
+		if err := common.UnmarshalJsonStr(data, &streamResponse); err != nil {
+			logger.LogError(c, "failed to unmarshal stream response: "+err.Error())
+			sr.Error(err)
+			return
+		}
+		sendResponsesStreamData(c, streamResponse, data)
+		switch streamResponse.Type {
+		case "response.completed", "response.done":
+			if streamResponse.Response != nil {
+				if streamResponse.Response.Usage != nil {
+					if streamResponse.Response.Usage.InputTokens != 0 {
+						usage.PromptTokens = streamResponse.Response.Usage.InputTokens
+					}
+					if streamResponse.Response.Usage.OutputTokens != 0 {
+						usage.CompletionTokens = streamResponse.Response.Usage.OutputTokens
+					}
+					if streamResponse.Response.Usage.TotalTokens != 0 {
+						usage.TotalTokens = streamResponse.Response.Usage.TotalTokens
+					}
+					if streamResponse.Response.Usage.InputTokensDetails != nil {
+						usage.PromptTokensDetails.CachedTokens = streamResponse.Response.Usage.InputTokensDetails.CachedTokens
+						usage.PromptTokensDetails.CacheWriteTokens = streamResponse.Response.Usage.InputTokensDetails.CacheWriteTokens
+					}
+				}
+				if !imageCommitted {
+					if relaycommon.IsNonBillableResponsesStatus(streamResponse.Response.Status) {
+						imageCounter.Reset()
+						imageCounter.Commit(info)
+						imageCommitted = true
+					} else {
+						for i := range streamResponse.Response.Output {
+							idx := i
+							imageCounter.Observe(&streamResponse.Response.Output[i], &idx)
+						}
+						imageCounter.Commit(info)
+						imageCommitted = true
+					}
+				}
+			} else if !imageCommitted {
+				imageCounter.Commit(info)
+				imageCommitted = true
+			}
+		case "response.failed", "response.incomplete", "response.cancelled", "response.canceled":
+			if !imageCommitted {
+				imageCounter.Reset()
+				imageCounter.Commit(info)
+				imageCommitted = true
+			}
+		case "response.output_text.delta":
+			// 处理输出文本
+			responseTextBuilder.WriteString(streamResponse.Delta)
+		case dto.ResponsesOutputTypeItemDone:
+			if streamResponse.Item != nil {
+				switch streamResponse.Item.Type {
+				case dto.BuildInCallWebSearchCall:
+					info.CountBillableToolCall(dto.BuildInCallWebSearchCall, "")
+				case dto.BuildInCallFileSearchCall:
+					info.CountBillableToolCall(dto.BuildInCallFileSearchCall, "")
+				case dto.BuildInCallFunctionCall:
+					info.CountBillableToolCall(dto.BuildInCallFunctionCall, streamResponse.Item.Name)
+				case dto.ResponsesOutputTypeImageGenerationCall:
+					if !imageCommitted {
+						imageCounter.Observe(streamResponse.Item, streamResponse.OutputIndex)
+					}
+				}
+			}
+		}
+	})
+
+	if usage.CompletionTokens == 0 {
+		// 计算输出文本的 token 数量
+		tempStr := responseTextBuilder.String()
+		if len(tempStr) > 0 {
+			// 非正常结束，使用输出文本的 token 数量
+			completionTokens := service.CountTextToken(tempStr, info.UpstreamModelName)
+			usage.CompletionTokens = completionTokens
+		}
+	}
+
+	if usage.PromptTokens == 0 && usage.CompletionTokens != 0 {
+		usage.PromptTokens = info.GetEstimatePromptTokens()
+	}
+
+	usage.TotalTokens = usage.PromptTokens + usage.CompletionTokens
+
+	return usage, nil
+}
@@ -0,0 +1,267 @@
+package openai
+
+import (
+	"bytes"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/QuantumNous/new-api/setting/operation_setting"
+	"github.com/gin-gonic/gin"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestOaiResponsesHandlerCountsOutputCallsNotDeclarations(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	operation_setting.SetToolPriceForTest("priced_fn", 5.0)
+	t.Cleanup(func() {
+		operation_setting.DeleteToolPriceForTest("priced_fn")
+	})
+
+	body, err := common.Marshal(dto.OpenAIResponsesResponse{
+		Tools: []map[string]any{
+			{"type": "web_search_preview"},
+			{"type": "file_search"},
+		},
+		Output: []dto.ResponsesOutput{
+			{Type: dto.BuildInCallWebSearchCall},
+			{Type: dto.BuildInCallWebSearchCall},
+			{Type: dto.BuildInCallFunctionCall, Name: "priced_fn"},
+			{Type: dto.BuildInCallFunctionCall, Name: "unpriced_fn"},
+		},
+		Usage: &dto.Usage{InputTokens: 1, OutputTokens: 1, TotalTokens: 2},
+	})
+	require.NoError(t, err)
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
+
+	info := &relaycommon.RelayInfo{
+		OriginModelName: "gpt-5.1",
+		ResponsesUsageInfo: &relaycommon.ResponsesUsageInfo{
+			BuiltInTools: map[string]*relaycommon.BuildInToolInfo{
+				dto.BuildInToolWebSearchPreview: {ToolName: dto.BuildInToolWebSearchPreview, CallCount: 0},
+				dto.BuildInToolFileSearch:       {ToolName: dto.BuildInToolFileSearch, CallCount: 0},
+			},
+		},
+	}
+	resp := &http.Response{
+		StatusCode: http.StatusOK,
+		Body:       io.NopCloser(bytes.NewReader(body)),
+		Header:     http.Header{"Content-Type": []string{"application/json"}},
+	}
+
+	usage, apiErr := OaiResponsesHandler(c, info, resp)
+	require.Nil(t, apiErr)
+	require.NotNil(t, usage)
+	assert.Equal(t, 2, info.ResponsesUsageInfo.BuiltInTools[dto.BuildInToolWebSearchPreview].CallCount)
+	assert.Equal(t, 0, info.ResponsesUsageInfo.BuiltInTools[dto.BuildInToolFileSearch].CallCount)
+	require.Contains(t, info.ResponsesUsageInfo.BuiltInTools, "priced_fn")
+	assert.Equal(t, 1, info.ResponsesUsageInfo.BuiltInTools["priced_fn"].CallCount)
+	assert.NotContains(t, info.ResponsesUsageInfo.BuiltInTools, "unpriced_fn")
+}
+
+func TestOaiResponsesHandlerDeclaredToolsWithoutOutputCountZero(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	body, err := common.Marshal(dto.OpenAIResponsesResponse{
+		Tools: []map[string]any{
+			{"type": "web_search_preview"},
+			{"type": "file_search"},
+		},
+		Output: []dto.ResponsesOutput{
+			{Type: "message", Role: "assistant"},
+		},
+		Usage: &dto.Usage{InputTokens: 1, OutputTokens: 1, TotalTokens: 2},
+	})
+	require.NoError(t, err)
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
+
+	info := &relaycommon.RelayInfo{
+		OriginModelName: "gpt-5.1",
+		ResponsesUsageInfo: &relaycommon.ResponsesUsageInfo{
+			BuiltInTools: map[string]*relaycommon.BuildInToolInfo{
+				dto.BuildInToolWebSearchPreview: {ToolName: dto.BuildInToolWebSearchPreview, CallCount: 0},
+				dto.BuildInToolFileSearch:       {ToolName: dto.BuildInToolFileSearch, CallCount: 0},
+			},
+		},
+	}
+	resp := &http.Response{
+		StatusCode: http.StatusOK,
+		Body:       io.NopCloser(bytes.NewReader(body)),
+		Header:     http.Header{"Content-Type": []string{"application/json"}},
+	}
+
+	_, apiErr := OaiResponsesHandler(c, info, resp)
+	require.Nil(t, apiErr)
+	assert.Equal(t, 0, info.ResponsesUsageInfo.BuiltInTools[dto.BuildInToolWebSearchPreview].CallCount)
+	assert.Equal(t, 0, info.ResponsesUsageInfo.BuiltInTools[dto.BuildInToolFileSearch].CallCount)
+}
+
+func TestOaiResponsesHandlerCountsCompletedImageGenerationOutputs(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	body, err := common.Marshal(dto.OpenAIResponsesResponse{
+		Status: []byte(`"completed"`),
+		Output: []dto.ResponsesOutput{
+			{
+				Type:   dto.ResponsesOutputTypeImageGenerationCall,
+				ID:     "img_1",
+				Status: "completed",
+				Result: "base64-a",
+			},
+			{
+				Type:   dto.ResponsesOutputTypeImageGenerationCall,
+				ID:     "img_2",
+				Status: "completed",
+				Result: "base64-b",
+			},
+			{
+				Type:   dto.ResponsesOutputTypeImageGenerationCall,
+				ID:     "img_empty",
+				Status: "completed",
+				Result: "",
+			},
+		},
+		Usage: &dto.Usage{InputTokens: 1, OutputTokens: 1, TotalTokens: 2},
+	})
+	require.NoError(t, err)
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
+	info := &relaycommon.RelayInfo{OriginModelName: "gpt-5.1"}
+	resp := &http.Response{
+		StatusCode: http.StatusOK,
+		Body:       io.NopCloser(bytes.NewReader(body)),
+		Header:     http.Header{"Content-Type": []string{"application/json"}},
+	}
+
+	_, apiErr := OaiResponsesHandler(c, info, resp)
+	require.Nil(t, apiErr)
+	require.Contains(t, info.ResponsesUsageInfo.BuiltInTools, dto.BuildInToolImageGeneration)
+	assert.Equal(t, 2, info.ResponsesUsageInfo.BuiltInTools[dto.BuildInToolImageGeneration].CallCount)
+	assert.False(t, c.GetBool("image_generation_call"))
+}
+
+func TestOaiResponsesHandlerIncompleteStatusCommitsZeroImageGeneration(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	body, err := common.Marshal(dto.OpenAIResponsesResponse{
+		Status: []byte(`"incomplete"`),
+		Output: []dto.ResponsesOutput{
+			{
+				Type:   dto.ResponsesOutputTypeImageGenerationCall,
+				ID:     "img_1",
+				Status: "completed",
+				Result: "base64-a",
+			},
+		},
+		Usage: &dto.Usage{InputTokens: 1, OutputTokens: 1, TotalTokens: 2},
+	})
+	require.NoError(t, err)
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
+	info := &relaycommon.RelayInfo{
+		OriginModelName: "gpt-5.1",
+		ResponsesUsageInfo: &relaycommon.ResponsesUsageInfo{
+			BuiltInTools: map[string]*relaycommon.BuildInToolInfo{
+				dto.BuildInToolImageGeneration: {ToolName: dto.BuildInToolImageGeneration, CallCount: 0},
+			},
+		},
+	}
+	resp := &http.Response{
+		StatusCode: http.StatusOK,
+		Body:       io.NopCloser(bytes.NewReader(body)),
+		Header:     http.Header{"Content-Type": []string{"application/json"}},
+	}
+
+	_, apiErr := OaiResponsesHandler(c, info, resp)
+	require.Nil(t, apiErr)
+	assert.Equal(t, 0, info.ResponsesUsageInfo.BuiltInTools[dto.BuildInToolImageGeneration].CallCount)
+}
+
+func runResponsesImageBillingStream(t *testing.T, events ...string) *relaycommon.RelayInfo {
+	t.Helper()
+	gin.SetMode(gin.TestMode)
+	oldTimeout := constant.StreamingTimeout
+	constant.StreamingTimeout = 30
+	t.Cleanup(func() {
+		constant.StreamingTimeout = oldTimeout
+	})
+
+	var body strings.Builder
+	for _, event := range events {
+		body.WriteString("data: ")
+		body.WriteString(event)
+		body.WriteString("\n\n")
+	}
+	body.WriteString("data: [DONE]\n\n")
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
+	c.Set(common.RequestIdKey, "responses-image-billing-test")
+	info := &relaycommon.RelayInfo{
+		OriginModelName: "gpt-5.1",
+		DisablePing:     true,
+		ChannelMeta: &relaycommon.ChannelMeta{
+			UpstreamModelName: "gpt-5.1",
+		},
+	}
+	resp := &http.Response{
+		StatusCode: http.StatusOK,
+		Body:       io.NopCloser(strings.NewReader(body.String())),
+		Header:     http.Header{"Content-Type": []string{"text/event-stream"}},
+	}
+
+	_, apiErr := OaiResponsesStreamHandler(c, info, resp)
+	require.Nil(t, apiErr)
+	require.NotNil(t, info.ResponsesUsageInfo)
+	require.Contains(t, info.ResponsesUsageInfo.BuiltInTools, dto.BuildInToolImageGeneration)
+	return info
+}
+
+func TestOaiResponsesStreamHandlerDeduplicatesCompletedImageOutput(t *testing.T) {
+	item := `{"type":"image_generation_call","id":"img_1","call_id":"call_1","status":"completed","result":"base64-a"}`
+	info := runResponsesImageBillingStream(
+		t,
+		`{"type":"response.output_item.done","output_index":0,"item":`+item+`}`,
+		`{"type":"response.completed","response":{"status":"completed","output":[`+item+`],"usage":{"input_tokens":1,"output_tokens":1,"total_tokens":2}}}`,
+	)
+
+	assert.Equal(t, 1, info.ResponsesUsageInfo.BuiltInTools[dto.BuildInToolImageGeneration].CallCount)
+}
+
+func TestOaiResponsesStreamHandlerDiscardsImageOutputOnIncomplete(t *testing.T) {
+	info := runResponsesImageBillingStream(
+		t,
+		`{"type":"response.output_item.done","output_index":0,"item":{"type":"image_generation_call","id":"img_1","status":"completed","result":"base64-a"}}`,
+		`{"type":"response.incomplete","response":{"status":"incomplete"}}`,
+	)
+
+	assert.Equal(t, 0, info.ResponsesUsageInfo.BuiltInTools[dto.BuildInToolImageGeneration].CallCount)
+}
+
+func TestOaiResponsesStreamHandlerDoesNotCountPartialImageEvent(t *testing.T) {
+	info := runResponsesImageBillingStream(
+		t,
+		`{"type":"response.image_generation_call.partial_image","output_index":0,"partial_image_b64":"partial-bytes"}`,
+		`{"type":"response.completed","response":{"status":"completed","output":[]}}`,
+	)
+
+	assert.Equal(t, 0, info.ResponsesUsageInfo.BuiltInTools[dto.BuildInToolImageGeneration].CallCount)
+}
@@ -0,0 +1,45 @@
+package openai
+
+import (
+	"io"
+	"net/http"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/QuantumNous/new-api/relaykit/types"
+	"github.com/QuantumNous/new-api/service"
+
+	"github.com/gin-gonic/gin"
+)
+
+func OaiResponsesCompactionHandler(c *gin.Context, resp *http.Response) (*dto.Usage, *types.NewAPIError) {
+	defer service.CloseResponseBodyGracefully(resp)
+
+	responseBody, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeReadResponseBodyFailed, http.StatusInternalServerError)
+	}
+
+	var compactResp dto.OpenAIResponsesCompactionResponse
+	if err := common.Unmarshal(responseBody, &compactResp); err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+	}
+	if oaiError := compactResp.GetOpenAIError(); oaiError != nil && oaiError.Type != "" {
+		return nil, types.WithOpenAIError(*oaiError, resp.StatusCode)
+	}
+
+	service.IOCopyBytesGracefully(c, resp, responseBody)
+
+	usage := dto.Usage{}
+	if compactResp.Usage != nil {
+		usage.PromptTokens = compactResp.Usage.InputTokens
+		usage.CompletionTokens = compactResp.Usage.OutputTokens
+		usage.TotalTokens = compactResp.Usage.TotalTokens
+		if compactResp.Usage.InputTokensDetails != nil {
+			usage.PromptTokensDetails.CachedTokens = compactResp.Usage.InputTokensDetails.CachedTokens
+			usage.PromptTokensDetails.CacheWriteTokens = compactResp.Usage.InputTokensDetails.CacheWriteTokens
+		}
+	}
+
+	return &usage, nil
+}
@@ -0,0 +1,158 @@
+package openai
+
+import (
+	"fmt"
+	"io"
+	"net/http"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/logger"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/QuantumNous/new-api/relay/helper"
+	"github.com/QuantumNous/new-api/relaykit/dto"
+	"github.com/QuantumNous/new-api/relaykit/relayconvert"
+	"github.com/QuantumNous/new-api/relaykit/types"
+	"github.com/QuantumNous/new-api/service"
+	"github.com/gin-gonic/gin"
+)
+
+func OaiChatToResponsesHandler(c *gin.Context, info *relaycommon.RelayInfo, resp *http.Response) (*dto.Usage, *types.NewAPIError) {
+	if resp == nil || resp.Body == nil {
+		return nil, types.NewOpenAIError(fmt.Errorf("invalid response"), types.ErrorCodeBadResponse, http.StatusInternalServerError)
+	}
+	defer service.CloseResponseBodyGracefully(resp)
+
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeReadResponseBodyFailed, http.StatusInternalServerError)
+	}
+
+	var chatResp dto.OpenAITextResponse
+	if err := common.Unmarshal(body, &chatResp); err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+	}
+	if oaiError := chatResp.GetOpenAIError(); oaiError != nil && oaiError.Type != "" {
+		return nil, types.WithOpenAIError(*oaiError, resp.StatusCode)
+	}
+
+	if responseID := helper.GetResponseID(c); responseID != "" {
+		chatResp.Id = responseID
+	}
+	convertResult, err := relayconvert.ConvertResponse(c, info, types.RelayFormatOpenAIResponses, &chatResp)
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+	}
+	responsesResp, ok := convertResult.Value.(*dto.OpenAIResponsesResponse)
+	if !ok {
+		return nil, types.NewOpenAIError(fmt.Errorf("expected OpenAI responses response, got %T", convertResult.Value), types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
+	}
+	usage := convertResult.Usage
+	if usage == nil || usage.TotalTokens == 0 {
+		text := service.ExtractOutputTextFromResponses(responsesResp)
+		usage = service.ResponseText2Usage(c, text, info.UpstreamModelName, info.GetEstimatePromptTokens())
+		responsesResp.Usage = relayconvert.UsageFromChatUsage(usage)
+	}
+
+	responseBody, err := common.Marshal(responsesResp)
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeJsonMarshalFailed, http.StatusInternalServerError)
+	}
+
+	service.IOCopyBytesGracefully(c, resp, responseBody)
+	return usage, nil
+}
+
+func OaiChatToResponsesStreamHandler(c *gin.Context, info *relaycommon.RelayInfo, resp *http.Response) (*dto.Usage, *types.NewAPIError) {
+	if resp == nil || resp.Body == nil {
+		return nil, types.NewOpenAIError(fmt.Errorf("invalid response"), types.ErrorCodeBadResponse, http.StatusInternalServerError)
+	}
+	defer service.CloseResponseBodyGracefully(resp)
+
+	responseID := helper.GetResponseID(c)
+	state, err := relayconvert.NewResponseStreamState(types.RelayFormatOpenAI, types.RelayFormatOpenAIResponses, relayconvert.ResponseStreamOptions{
+		ID:    responseID,
+		Model: info.UpstreamModelName,
+	})
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponse, http.StatusInternalServerError)
+	}
+	streamErr := (*types.NewAPIError)(nil)
+
+	sendEvent := func(event relayconvert.ChatToResponsesStreamEvent) bool {
+		data, err := common.Marshal(event.Payload)
+		if err != nil {
+			streamErr = types.NewOpenAIError(err, types.ErrorCodeJsonMarshalFailed, http.StatusInternalServerError)
+			return false
+		}
+		helper.ResponseChunkData(c, dto.ResponsesStreamResponse{Type: event.Type}, string(data))
+		return true
+	}
+
+	helper.StreamScannerHandler(c, resp, info, func(data string, sr *helper.StreamResult) {
+		if streamErr != nil {
+			sr.Stop(streamErr)
+			return
+		}
+
+		var errorResp dto.OpenAITextResponse
+		if err := common.UnmarshalJsonStr(data, &errorResp); err == nil {
+			if oaiError := errorResp.GetOpenAIError(); oaiError != nil && oaiError.Type != "" {
+				streamErr = types.WithOpenAIError(*oaiError, resp.StatusCode)
+				sr.Stop(streamErr)
+				return
+			}
+		}
+
+		var chunk dto.ChatCompletionsStreamResponse
+		if err := common.UnmarshalJsonStr(data, &chunk); err != nil {
+			logger.LogError(c, "failed to unmarshal chat stream response: "+err.Error())
+			sr.Error(err)
+			return
+		}
+
+		results, err := relayconvert.ConvertStreamResponseChunk(c, info, state, &chunk)
+		if err != nil {
+			streamErr = types.NewOpenAIError(err, types.ErrorCodeBadResponse, http.StatusInternalServerError)
+			sr.Stop(streamErr)
+			return
+		}
+		for _, result := range results {
+			event, ok := result.Value.(relayconvert.ChatToResponsesStreamEvent)
+			if !ok {
+				streamErr = types.NewOpenAIError(fmt.Errorf("expected OAI responses stream event, got %T", result.Value), types.ErrorCodeBadResponse, http.StatusInternalServerError)
+				sr.Stop(streamErr)
+				return
+			}
+			if !sendEvent(event) {
+				sr.Stop(streamErr)
+				return
+			}
+		}
+	})
+
+	if streamErr != nil {
+		return nil, streamErr
+	}
+
+	usage := state.Usage()
+	if usage == nil || usage.TotalTokens == 0 {
+		usage = service.ResponseText2Usage(c, state.UsageText(), info.UpstreamModelName, info.GetEstimatePromptTokens())
+		state.SetUsage(usage)
+	}
+
+	finalResults, err := relayconvert.FinalizeStreamResponse(c, info, state)
+	if err != nil {
+		return nil, types.NewOpenAIError(err, types.ErrorCodeBadResponse, http.StatusInternalServerError)
+	}
+	for _, result := range finalResults {
+		event, ok := result.Value.(relayconvert.ChatToResponsesStreamEvent)
+		if !ok {
+			return nil, types.NewOpenAIError(fmt.Errorf("expected OAI responses stream event, got %T", result.Value), types.ErrorCodeBadResponse, http.StatusInternalServerError)
+		}
+		if !sendEvent(event) {
+			return nil, streamErr
+		}
+	}
+
+	return usage, nil
+}
@@ -0,0 +1,27 @@
+package openai
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestCollectStreamFunctionCallNamesDedupesSameIndex(t *testing.T) {
+	seen := make(map[string]struct{})
+	var names []string
+
+	chunks := []string{
+		`{"choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"id":"c1","type":"function","function":{"name":"get_weather","arguments":""}}]}}]}`,
+		`{"choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"{\"q\":"}}]}}]}`,
+		`{"choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\"x\"}"}}]}}]}`,
+		`{"choices":[{"index":0,"delta":{"tool_calls":[{"index":1,"id":"c2","type":"function","function":{"name":"get_time","arguments":""}}]}}]}`,
+		`{"choices":[{"index":0,"delta":{"tool_calls":[{"index":1,"function":{"arguments":"{}"}}]}}]}`,
+	}
+	for _, chunk := range chunks {
+		collectStreamFunctionCallNames(chunk, seen, &names)
+	}
+
+	require.Len(t, names, 2)
+	assert.Equal(t, []string{"get_weather", "get_time"}, names)
+}
@@ -0,0 +1,133 @@
+package openai
+
+import (
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/QuantumNous/new-api/relaykit/dto"
+)
+
+func applyUsagePostProcessing(info *relaycommon.RelayInfo, usage *dto.Usage, responseBody []byte) {
+	if info == nil || usage == nil {
+		return
+	}
+
+	switch info.ChannelType {
+	case constant.ChannelTypeDeepSeek:
+		if usage.PromptTokensDetails.CachedTokens == 0 && usage.PromptCacheHitTokens != 0 {
+			usage.PromptTokensDetails.CachedTokens = usage.PromptCacheHitTokens
+		}
+	case constant.ChannelTypeZhipu_v4:
+		// 智普的cached_tokens在标准位置: usage.prompt_tokens_details.cached_tokens
+		if usage.PromptTokensDetails.CachedTokens == 0 {
+			if usage.InputTokensDetails != nil && usage.InputTokensDetails.CachedTokens > 0 {
+				usage.PromptTokensDetails.CachedTokens = usage.InputTokensDetails.CachedTokens
+			} else if cachedTokens, ok := extractCachedTokensFromBody(responseBody); ok {
+				usage.PromptTokensDetails.CachedTokens = cachedTokens
+			} else if usage.PromptCacheHitTokens > 0 {
+				usage.PromptTokensDetails.CachedTokens = usage.PromptCacheHitTokens
+			}
+		}
+	case constant.ChannelTypeMoonshot:
+		// Moonshot的cached_tokens在非标准位置: choices[].usage.cached_tokens
+		if usage.PromptTokensDetails.CachedTokens == 0 {
+			if usage.InputTokensDetails != nil && usage.InputTokensDetails.CachedTokens > 0 {
+				usage.PromptTokensDetails.CachedTokens = usage.InputTokensDetails.CachedTokens
+			} else if cachedTokens, ok := extractMoonshotCachedTokensFromBody(responseBody); ok {
+				usage.PromptTokensDetails.CachedTokens = cachedTokens
+			} else if cachedTokens, ok := extractCachedTokensFromBody(responseBody); ok {
+				usage.PromptTokensDetails.CachedTokens = cachedTokens
+			} else if usage.PromptCacheHitTokens > 0 {
+				usage.PromptTokensDetails.CachedTokens = usage.PromptCacheHitTokens
+			}
+		}
+	case constant.ChannelTypeOpenAI:
+		if usage.PromptTokensDetails.CachedTokens == 0 {
+			if cachedTokens, ok := extractLlamaCachedTokensFromBody(responseBody); ok {
+				usage.PromptTokensDetails.CachedTokens = cachedTokens
+			}
+		}
+	}
+}
+
+func extractCachedTokensFromBody(body []byte) (int, bool) {
+	if len(body) == 0 {
+		return 0, false
+	}
+
+	var payload struct {
+		Usage struct {
+			PromptTokensDetails struct {
+				CachedTokens *int `json:"cached_tokens"`
+			} `json:"prompt_tokens_details"`
+			CachedTokens         *int `json:"cached_tokens"`
+			PromptCacheHitTokens *int `json:"prompt_cache_hit_tokens"`
+		} `json:"usage"`
+	}
+
+	if err := common.Unmarshal(body, &payload); err != nil {
+		return 0, false
+	}
+
+	if payload.Usage.PromptTokensDetails.CachedTokens != nil {
+		return *payload.Usage.PromptTokensDetails.CachedTokens, true
+	}
+	if payload.Usage.CachedTokens != nil {
+		return *payload.Usage.CachedTokens, true
+	}
+	if payload.Usage.PromptCacheHitTokens != nil {
+		return *payload.Usage.PromptCacheHitTokens, true
+	}
+	return 0, false
+}
+
+// extractMoonshotCachedTokensFromBody 从Moonshot的非标准位置提取cached_tokens
+// Moonshot的流式响应格式: {"choices":[{"usage":{"cached_tokens":111}}]}
+func extractMoonshotCachedTokensFromBody(body []byte) (int, bool) {
+	if len(body) == 0 {
+		return 0, false
+	}
+
+	var payload struct {
+		Choices []struct {
+			Usage struct {
+				CachedTokens *int `json:"cached_tokens"`
+			} `json:"usage"`
+		} `json:"choices"`
+	}
+
+	if err := common.Unmarshal(body, &payload); err != nil {
+		return 0, false
+	}
+
+	// 遍历choices查找cached_tokens
+	for _, choice := range payload.Choices {
+		if choice.Usage.CachedTokens != nil && *choice.Usage.CachedTokens > 0 {
+			return *choice.Usage.CachedTokens, true
+		}
+	}
+
+	return 0, false
+}
+
+// extractLlamaCachedTokensFromBody 从llama.cpp的非标准位置提取cache_n
+func extractLlamaCachedTokensFromBody(body []byte) (int, bool) {
+	if len(body) == 0 {
+		return 0, false
+	}
+
+	var payload struct {
+		Timings struct {
+			CachedTokens *int `json:"cache_n"`
+		} `json:"timings"`
+	}
+
+	if err := common.Unmarshal(body, &payload); err != nil {
+		return 0, false
+	}
+
+	if payload.Timings.CachedTokens == nil {
+		return 0, false
+	}
+	return *payload.Timings.CachedTokens, true
+}