Merge branch 'test' into release

backend/internal/handler/admin/account_handler_passthrough_test.go

+package admin
+
+import (
+	"bytes"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+	"github.com/stretchr/testify/require"
+)
+
+func TestAccountHandler_Create_AnthropicAPIKeyPassthroughExtraForwarded(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	adminSvc := newStubAdminService()
+	handler := NewAccountHandler(
+		adminSvc,
+		nil,
+		nil,
+		nil,
+		nil,
+		nil,
+		nil,
+		nil,
+		nil,
+		nil,
+		nil,
+		nil,
+	)
+
+	router := gin.New()
+	router.POST("/api/v1/admin/accounts", handler.Create)
+
+	body := map[string]any{
+		"name":     "anthropic-key-1",
+		"platform": "anthropic",
+		"type":     "apikey",
+		"credentials": map[string]any{
+			"api_key":  "sk-ant-xxx",
+			"base_url": "https://api.anthropic.com",
+		},
+		"extra": map[string]any{
+			"anthropic_passthrough": true,
+		},
+		"concurrency": 1,
+		"priority":    1,
+	}
+	raw, err := json.Marshal(body)
+	require.NoError(t, err)
+
+	rec := httptest.NewRecorder()
+	req := httptest.NewRequest(http.MethodPost, "/api/v1/admin/accounts", bytes.NewReader(raw))
+	req.Header.Set("Content-Type", "application/json")
+	router.ServeHTTP(rec, req)
+
+	require.Equal(t, http.StatusOK, rec.Code)
+	require.Len(t, adminSvc.createdAccounts, 1)
+
+	created := adminSvc.createdAccounts[0]
+	require.Equal(t, "anthropic", created.Platform)
+	require.Equal(t, "apikey", created.Type)
+	require.NotNil(t, created.Extra)
+	require.Equal(t, true, created.Extra["anthropic_passthrough"])
+}
+

backend/internal/server/routes/gateway.go

@@ -43,12 +43,12 @@ func RegisterGatewayRoutes(
 		gateway.GET("/usage", h.Gateway.Usage)
 		// OpenAI Responses API
 		gateway.POST("/responses", h.OpenAIGateway.Responses)
-		// 明确阻止旧入口误用到 Sora，避免客户端把 OpenAI Chat Completions 当作 Sora 入口
+		// 明确阻止旧协议入口：OpenAI 仅支持 Responses API，避免客户端误解为会自动路由到其它平台。
 		gateway.POST("/chat/completions", func(c *gin.Context) {
 			c.JSON(http.StatusBadRequest, gin.H{
 				"error": gin.H{
 					"type":    "invalid_request_error",
-					"message": "For Sora, use /sora/v1/chat/completions. OpenAI should use /v1/responses.",
+					"message": "Unsupported legacy protocol: /v1/chat/completions is not supported. Please use /v1/responses.",
 				},
 			})
 		})

backend/internal/service/account.go

@@ -719,6 +719,17 @@ func (a *Account) IsOpenAIOAuthPassthroughEnabled() bool {
 	return a != nil && a.IsOpenAIOAuth() && a.IsOpenAIPassthroughEnabled()
 }
 
+// IsAnthropicAPIKeyPassthroughEnabled 返回 Anthropic API Key 账号是否启用“自动透传（仅替换认证）”。
+// 字段：accounts.extra.anthropic_passthrough。
+// 字段缺失或类型不正确时，按 false（关闭）处理。
+func (a *Account) IsAnthropicAPIKeyPassthroughEnabled() bool {
+	if a == nil || a.Platform != PlatformAnthropic || a.Type != AccountTypeAPIKey || a.Extra == nil {
+		return false
+	}
+	enabled, ok := a.Extra["anthropic_passthrough"].(bool)
+	return ok && enabled
+}
+
 // IsCodexCLIOnlyEnabled 返回 OpenAI OAuth 账号是否启用“仅允许 Codex 官方客户端”。
 // 字段：accounts.extra.codex_cli_only。
 // 字段缺失或类型不正确时，按 false（关闭）处理。

backend/internal/service/account_anthropic_passthrough_test.go

+package service
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+func TestAccount_IsAnthropicAPIKeyPassthroughEnabled(t *testing.T) {
+	t.Run("Anthropic API Key 开启", func(t *testing.T) {
+		account := &Account{
+			Platform: PlatformAnthropic,
+			Type:     AccountTypeAPIKey,
+			Extra: map[string]any{
+				"anthropic_passthrough": true,
+			},
+		}
+		require.True(t, account.IsAnthropicAPIKeyPassthroughEnabled())
+	})
+
+	t.Run("Anthropic API Key 关闭", func(t *testing.T) {
+		account := &Account{
+			Platform: PlatformAnthropic,
+			Type:     AccountTypeAPIKey,
+			Extra: map[string]any{
+				"anthropic_passthrough": false,
+			},
+		}
+		require.False(t, account.IsAnthropicAPIKeyPassthroughEnabled())
+	})
+
+	t.Run("字段类型非法默认关闭", func(t *testing.T) {
+		account := &Account{
+			Platform: PlatformAnthropic,
+			Type:     AccountTypeAPIKey,
+			Extra: map[string]any{
+				"anthropic_passthrough": "true",
+			},
+		}
+		require.False(t, account.IsAnthropicAPIKeyPassthroughEnabled())
+	})
+
+	t.Run("非 Anthropic API Key 账号始终关闭", func(t *testing.T) {
+		oauth := &Account{
+			Platform: PlatformAnthropic,
+			Type:     AccountTypeOAuth,
+			Extra: map[string]any{
+				"anthropic_passthrough": true,
+			},
+		}
+		require.False(t, oauth.IsAnthropicAPIKeyPassthroughEnabled())
+
+		openai := &Account{
+			Platform: PlatformOpenAI,
+			Type:     AccountTypeAPIKey,
+			Extra: map[string]any{
+				"anthropic_passthrough": true,
+			},
+		}
+		require.False(t, openai.IsAnthropicAPIKeyPassthroughEnabled())
+	})
+}

backend/internal/service/gateway_anthropic_apikey_passthrough_benchmark_test.go

+package service
+
+import "testing"
+
+func BenchmarkGatewayService_ParseSSEUsage_MessageStart(b *testing.B) {
+	svc := &GatewayService{}
+	data := `{"type":"message_start","message":{"usage":{"input_tokens":123,"cache_creation_input_tokens":45,"cache_read_input_tokens":6,"cached_tokens":6,"cache_creation":{"ephemeral_5m_input_tokens":20,"ephemeral_1h_input_tokens":25}}}}`
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		usage := &ClaudeUsage{}
+		svc.parseSSEUsage(data, usage)
+	}
+}
+
+func BenchmarkGatewayService_ParseSSEUsagePassthrough_MessageStart(b *testing.B) {
+	svc := &GatewayService{}
+	data := `{"type":"message_start","message":{"usage":{"input_tokens":123,"cache_creation_input_tokens":45,"cache_read_input_tokens":6,"cached_tokens":6,"cache_creation":{"ephemeral_5m_input_tokens":20,"ephemeral_1h_input_tokens":25}}}}`
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		usage := &ClaudeUsage{}
+		svc.parseSSEUsagePassthrough(data, usage)
+	}
+}
+
+func BenchmarkGatewayService_ParseSSEUsage_MessageDelta(b *testing.B) {
+	svc := &GatewayService{}
+	data := `{"type":"message_delta","usage":{"output_tokens":456,"cache_creation_input_tokens":30,"cache_read_input_tokens":7,"cached_tokens":7,"cache_creation":{"ephemeral_5m_input_tokens":10,"ephemeral_1h_input_tokens":20}}}`
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		usage := &ClaudeUsage{}
+		svc.parseSSEUsage(data, usage)
+	}
+}
+
+func BenchmarkGatewayService_ParseSSEUsagePassthrough_MessageDelta(b *testing.B) {
+	svc := &GatewayService{}
+	data := `{"type":"message_delta","usage":{"output_tokens":456,"cache_creation_input_tokens":30,"cache_read_input_tokens":7,"cached_tokens":7,"cache_creation":{"ephemeral_5m_input_tokens":10,"ephemeral_1h_input_tokens":20}}}`
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		usage := &ClaudeUsage{}
+		svc.parseSSEUsagePassthrough(data, usage)
+	}
+}
+
+func BenchmarkParseClaudeUsageFromResponseBody(b *testing.B) {
+	body := []byte(`{"id":"msg_123","type":"message","usage":{"input_tokens":123,"output_tokens":456,"cache_creation_input_tokens":45,"cache_read_input_tokens":6,"cached_tokens":6,"cache_creation":{"ephemeral_5m_input_tokens":20,"ephemeral_1h_input_tokens":25}}}`)
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		_ = parseClaudeUsageFromResponseBody(body)
+	}
+}

backend/internal/service/gateway_anthropic_apikey_passthrough_test.go

+package service
+
+import (
+	"bufio"
+	"bytes"
+	"context"
+	"errors"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/Wei-Shaw/sub2api/internal/config"
+	"github.com/Wei-Shaw/sub2api/internal/pkg/claude"
+	"github.com/gin-gonic/gin"
+	"github.com/stretchr/testify/require"
+	"github.com/tidwall/gjson"
+)
+
+type anthropicHTTPUpstreamRecorder struct {
+	lastReq  *http.Request
+	lastBody []byte
+	resp     *http.Response
+	err      error
+}
+
+func newAnthropicAPIKeyAccountForTest() *Account {
+	return &Account{
+		ID:          201,
+		Name:        "anthropic-apikey-pass-test",
+		Platform:    PlatformAnthropic,
+		Type:        AccountTypeAPIKey,
+		Concurrency: 1,
+		Credentials: map[string]any{
+			"api_key":  "upstream-anthropic-key",
+			"base_url": "https://api.anthropic.com",
+		},
+		Extra: map[string]any{
+			"anthropic_passthrough": true,
+		},
+		Status:      StatusActive,
+		Schedulable: true,
+	}
+}
+
+func (u *anthropicHTTPUpstreamRecorder) Do(req *http.Request, proxyURL string, accountID int64, accountConcurrency int) (*http.Response, error) {
+	u.lastReq = req
+	if req != nil && req.Body != nil {
+		b, _ := io.ReadAll(req.Body)
+		u.lastBody = b
+		_ = req.Body.Close()
+		req.Body = io.NopCloser(bytes.NewReader(b))
+	}
+	if u.err != nil {
+		return nil, u.err
+	}
+	return u.resp, nil
+}
+
+func (u *anthropicHTTPUpstreamRecorder) DoWithTLS(req *http.Request, proxyURL string, accountID int64, accountConcurrency int, enableTLSFingerprint bool) (*http.Response, error) {
+	return u.Do(req, proxyURL, accountID, accountConcurrency)
+}
+
+type streamReadCloser struct {
+	payload []byte
+	sent    bool
+	err     error
+}
+
+func (r *streamReadCloser) Read(p []byte) (int, error) {
+	if !r.sent {
+		r.sent = true
+		n := copy(p, r.payload)
+		return n, nil
+	}
+	if r.err != nil {
+		return 0, r.err
+	}
+	return 0, io.EOF
+}
+
+func (r *streamReadCloser) Close() error { return nil }
+
+type failWriteResponseWriter struct {
+	gin.ResponseWriter
+}
+
+func (w *failWriteResponseWriter) Write(data []byte) (int, error) {
+	return 0, errors.New("client disconnected")
+}
+
+func (w *failWriteResponseWriter) WriteString(_ string) (int, error) {
+	return 0, errors.New("client disconnected")
+}
+
+func TestGatewayService_AnthropicAPIKeyPassthrough_ForwardStreamPreservesBodyAndAuthReplacement(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages", nil)
+	c.Request.Header.Set("User-Agent", "claude-cli/1.0.0")
+	c.Request.Header.Set("Authorization", "Bearer inbound-token")
+	c.Request.Header.Set("X-Api-Key", "inbound-api-key")
+	c.Request.Header.Set("X-Goog-Api-Key", "inbound-goog-key")
+	c.Request.Header.Set("Cookie", "secret=1")
+	c.Request.Header.Set("Anthropic-Beta", "interleaved-thinking-2025-05-14")
+
+	body := []byte(`{"model":"claude-3-7-sonnet-20250219","stream":true,"system":[{"type":"text","text":"x-anthropic-billing-header keep"}],"messages":[{"role":"user","content":[{"type":"text","text":"hello"}]}]}`)
+	parsed := &ParsedRequest{
+		Body:   body,
+		Model:  "claude-3-7-sonnet-20250219",
+		Stream: true,
+	}
+
+	upstreamSSE := strings.Join([]string{
+		`data: {"type":"message_start","message":{"usage":{"input_tokens":9,"cached_tokens":7}}}`,
+		"",
+		`data: {"type":"message_delta","usage":{"output_tokens":3}}`,
+		"",
+		"data: [DONE]",
+		"",
+	}, "\n")
+	upstream := &anthropicHTTPUpstreamRecorder{
+		resp: &http.Response{
+			StatusCode: http.StatusOK,
+			Header: http.Header{
+				"Content-Type": []string{"text/event-stream"},
+				"x-request-id": []string{"rid-anthropic-pass"},
+				"Set-Cookie":   []string{"secret=upstream"},
+			},
+			Body: io.NopCloser(strings.NewReader(upstreamSSE)),
+		},
+	}
+
+	svc := &GatewayService{
+		cfg: &config.Config{
+			Gateway: config.GatewayConfig{
+				MaxLineSize: defaultMaxLineSize,
+			},
+		},
+		httpUpstream:        upstream,
+		rateLimitService:    &RateLimitService{},
+		deferredService:     &DeferredService{},
+		billingCacheService: nil,
+	}
+
+	account := &Account{
+		ID:          101,
+		Name:        "anthropic-apikey-pass",
+		Platform:    PlatformAnthropic,
+		Type:        AccountTypeAPIKey,
+		Concurrency: 1,
+		Credentials: map[string]any{
+			"api_key":       "upstream-anthropic-key",
+			"base_url":      "https://api.anthropic.com",
+			"model_mapping": map[string]any{"claude-3-7-sonnet-20250219": "claude-3-haiku-20240307"},
+		},
+		Extra: map[string]any{
+			"anthropic_passthrough": true,
+		},
+		Status:      StatusActive,
+		Schedulable: true,
+	}
+
+	result, err := svc.Forward(context.Background(), c, account, parsed)
+	require.NoError(t, err)
+	require.NotNil(t, result)
+	require.True(t, result.Stream)
+
+	require.Equal(t, body, upstream.lastBody, "透传模式不应改写上游请求体")
+	require.Equal(t, "claude-3-7-sonnet-20250219", gjson.GetBytes(upstream.lastBody, "model").String())
+
+	require.Equal(t, "upstream-anthropic-key", upstream.lastReq.Header.Get("x-api-key"))
+	require.Empty(t, upstream.lastReq.Header.Get("authorization"))
+	require.Empty(t, upstream.lastReq.Header.Get("x-goog-api-key"))
+	require.Empty(t, upstream.lastReq.Header.Get("cookie"))
+	require.Equal(t, "2023-06-01", upstream.lastReq.Header.Get("anthropic-version"))
+	require.Equal(t, "interleaved-thinking-2025-05-14", upstream.lastReq.Header.Get("anthropic-beta"))
+	require.Empty(t, upstream.lastReq.Header.Get("x-stainless-lang"), "API Key 透传不应注入 OAuth 指纹头")
+
+	require.Contains(t, rec.Body.String(), `"cached_tokens":7`)
+	require.NotContains(t, rec.Body.String(), `"cache_read_input_tokens":7`, "透传输出不应被网关改写")
+	require.Equal(t, 7, result.Usage.CacheReadInputTokens, "计费 usage 解析应保留 cached_tokens 兼容")
+	require.Empty(t, rec.Header().Get("Set-Cookie"), "响应头应经过安全过滤")
+	rawBody, ok := c.Get(OpsUpstreamRequestBodyKey)
+	require.True(t, ok)
+	bodyBytes, ok := rawBody.([]byte)
+	require.True(t, ok, "应以 []byte 形式缓存上游请求体，避免重复 string 拷贝")
+	require.Equal(t, body, bodyBytes)
+}
+
+func TestGatewayService_AnthropicAPIKeyPassthrough_ForwardCountTokensPreservesBody(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages/count_tokens", nil)
+	c.Request.Header.Set("Authorization", "Bearer inbound-token")
+	c.Request.Header.Set("X-Api-Key", "inbound-api-key")
+	c.Request.Header.Set("Cookie", "secret=1")
+
+	body := []byte(`{"model":"claude-3-5-sonnet-latest","messages":[{"role":"user","content":[{"type":"text","text":"hello"}]}],"thinking":{"type":"enabled"}}`)
+	parsed := &ParsedRequest{
+		Body:  body,
+		Model: "claude-3-5-sonnet-latest",
+	}
+
+	upstreamRespBody := `{"input_tokens":42}`
+	upstream := &anthropicHTTPUpstreamRecorder{
+		resp: &http.Response{
+			StatusCode: http.StatusOK,
+			Header: http.Header{
+				"Content-Type": []string{"application/json"},
+				"x-request-id": []string{"rid-count"},
+				"Set-Cookie":   []string{"secret=upstream"},
+			},
+			Body: io.NopCloser(strings.NewReader(upstreamRespBody)),
+		},
+	}
+
+	svc := &GatewayService{
+		cfg: &config.Config{
+			Gateway: config.GatewayConfig{
+				MaxLineSize: defaultMaxLineSize,
+			},
+		},
+		httpUpstream:     upstream,
+		rateLimitService: &RateLimitService{},
+	}
+
+	account := &Account{
+		ID:          102,
+		Name:        "anthropic-apikey-pass-count",
+		Platform:    PlatformAnthropic,
+		Type:        AccountTypeAPIKey,
+		Concurrency: 1,
+		Credentials: map[string]any{
+			"api_key":       "upstream-anthropic-key",
+			"base_url":      "https://api.anthropic.com",
+			"model_mapping": map[string]any{"claude-3-5-sonnet-latest": "claude-3-opus-20240229"},
+		},
+		Extra: map[string]any{
+			"anthropic_passthrough": true,
+		},
+		Status:      StatusActive,
+		Schedulable: true,
+	}
+
+	err := svc.ForwardCountTokens(context.Background(), c, account, parsed)
+	require.NoError(t, err)
+
+	require.Equal(t, body, upstream.lastBody, "count_tokens 透传模式不应改写请求体")
+	require.Equal(t, "claude-3-5-sonnet-latest", gjson.GetBytes(upstream.lastBody, "model").String())
+	require.Equal(t, "upstream-anthropic-key", upstream.lastReq.Header.Get("x-api-key"))
+	require.Empty(t, upstream.lastReq.Header.Get("authorization"))
+	require.Empty(t, upstream.lastReq.Header.Get("cookie"))
+	require.Equal(t, http.StatusOK, rec.Code)
+	require.JSONEq(t, upstreamRespBody, rec.Body.String())
+	require.Empty(t, rec.Header().Get("Set-Cookie"))
+}
+
+func TestGatewayService_AnthropicAPIKeyPassthrough_BuildRequestRejectsInvalidBaseURL(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages", nil)
+
+	svc := &GatewayService{
+		cfg: &config.Config{
+			Security: config.SecurityConfig{
+				URLAllowlist: config.URLAllowlistConfig{
+					Enabled: false,
+				},
+			},
+		},
+	}
+	account := &Account{
+		Platform: PlatformAnthropic,
+		Type:     AccountTypeAPIKey,
+		Credentials: map[string]any{
+			"api_key":  "k",
+			"base_url": "://invalid-url",
+		},
+	}
+
+	_, err := svc.buildUpstreamRequestAnthropicAPIKeyPassthrough(context.Background(), c, account, []byte(`{}`), "k")
+	require.Error(t, err)
+}
+
+func TestGatewayService_AnthropicOAuth_NotAffectedByAPIKeyPassthroughToggle(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages", nil)
+
+	svc := &GatewayService{
+		cfg: &config.Config{
+			Gateway: config.GatewayConfig{MaxLineSize: defaultMaxLineSize},
+		},
+	}
+	account := &Account{
+		Platform: PlatformAnthropic,
+		Type:     AccountTypeOAuth,
+		Extra: map[string]any{
+			"anthropic_passthrough": true,
+		},
+	}
+
+	require.False(t, account.IsAnthropicAPIKeyPassthroughEnabled())
+
+	req, err := svc.buildUpstreamRequest(context.Background(), c, account, []byte(`{"model":"claude-3-7-sonnet-20250219"}`), "oauth-token", "oauth", "claude-3-7-sonnet-20250219", true, false)
+	require.NoError(t, err)
+	require.Equal(t, "Bearer oauth-token", req.Header.Get("authorization"))
+	require.Contains(t, req.Header.Get("anthropic-beta"), claude.BetaOAuth, "OAuth 链路仍应按原逻辑补齐 oauth beta")
+}
+
+func TestGatewayService_AnthropicAPIKeyPassthrough_StreamingStillCollectsUsageAfterClientDisconnect(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	// Use a canceled context recorder to simulate client disconnect behavior.
+	req := httptest.NewRequest(http.MethodPost, "/v1/messages", nil)
+	ctx, cancel := context.WithCancel(req.Context())
+	cancel()
+	req = req.WithContext(ctx)
+
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = req
+
+	svc := &GatewayService{
+		cfg: &config.Config{
+			Gateway: config.GatewayConfig{
+				MaxLineSize: defaultMaxLineSize,
+			},
+		},
+		rateLimitService: &RateLimitService{},
+	}
+
+	resp := &http.Response{
+		StatusCode: http.StatusOK,
+		Header:     http.Header{"Content-Type": []string{"text/event-stream"}},
+		Body: io.NopCloser(strings.NewReader(strings.Join([]string{
+			`data: {"type":"message_start","message":{"usage":{"input_tokens":11}}}`,
+			"",
+			`data: {"type":"message_delta","usage":{"output_tokens":5}}`,
+			"",
+			"data: [DONE]",
+			"",
+		}, "\n"))),
+	}
+
+	result, err := svc.handleStreamingResponseAnthropicAPIKeyPassthrough(context.Background(), resp, c, &Account{ID: 1}, time.Now(), "claude-3-7-sonnet-20250219")
+	require.NoError(t, err)
+	require.NotNil(t, result)
+	require.NotNil(t, result.usage)
+	require.Equal(t, 11, result.usage.InputTokens)
+	require.Equal(t, 5, result.usage.OutputTokens)
+}
+
+func TestGatewayService_AnthropicAPIKeyPassthrough_ForwardDirect_NonStreamingSuccess(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages", nil)
+
+	body := []byte(`{"model":"claude-3-5-sonnet-latest","messages":[{"role":"user","content":[{"type":"text","text":"hello"}]}]}`)
+	upstreamJSON := `{"id":"msg_1","type":"message","usage":{"input_tokens":12,"output_tokens":7,"cache_creation":{"ephemeral_5m_input_tokens":2,"ephemeral_1h_input_tokens":3},"cached_tokens":4}}`
+	upstream := &anthropicHTTPUpstreamRecorder{
+		resp: &http.Response{
+			StatusCode: http.StatusOK,
+			Header: http.Header{
+				"Content-Type": []string{"application/json"},
+				"x-request-id": []string{"rid-nonstream"},
+			},
+			Body: io.NopCloser(strings.NewReader(upstreamJSON)),
+		},
+	}
+	svc := &GatewayService{
+		cfg:              &config.Config{},
+		httpUpstream:     upstream,
+		rateLimitService: &RateLimitService{},
+	}
+
+	result, err := svc.forwardAnthropicAPIKeyPassthrough(context.Background(), c, newAnthropicAPIKeyAccountForTest(), body, "claude-3-5-sonnet-latest", false, time.Now())
+	require.NoError(t, err)
+	require.NotNil(t, result)
+	require.Equal(t, 12, result.Usage.InputTokens)
+	require.Equal(t, 7, result.Usage.OutputTokens)
+	require.Equal(t, 5, result.Usage.CacheCreationInputTokens)
+	require.Equal(t, 4, result.Usage.CacheReadInputTokens)
+	require.Equal(t, upstreamJSON, rec.Body.String())
+}
+
+func TestGatewayService_AnthropicAPIKeyPassthrough_ForwardDirect_InvalidTokenType(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages", nil)
+
+	account := &Account{
+		ID:       202,
+		Name:     "anthropic-oauth",
+		Platform: PlatformAnthropic,
+		Type:     AccountTypeOAuth,
+		Credentials: map[string]any{
+			"access_token": "oauth-token",
+		},
+	}
+	svc := &GatewayService{}
+
+	result, err := svc.forwardAnthropicAPIKeyPassthrough(context.Background(), c, account, []byte(`{}`), "claude-3-5-sonnet-latest", false, time.Now())
+	require.Nil(t, result)
+	require.Error(t, err)
+	require.Contains(t, err.Error(), "requires apikey token")
+}
+
+func TestGatewayService_AnthropicAPIKeyPassthrough_ForwardDirect_UpstreamRequestError(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages", nil)
+
+	upstream := &anthropicHTTPUpstreamRecorder{
+		err: errors.New("dial tcp timeout"),
+	}
+	svc := &GatewayService{
+		cfg: &config.Config{
+			Security: config.SecurityConfig{
+				URLAllowlist: config.URLAllowlistConfig{Enabled: false},
+			},
+		},
+		httpUpstream: upstream,
+	}
+	account := newAnthropicAPIKeyAccountForTest()
+
+	result, err := svc.forwardAnthropicAPIKeyPassthrough(context.Background(), c, account, []byte(`{"model":"x"}`), "x", false, time.Now())
+	require.Nil(t, result)
+	require.Error(t, err)
+	require.Contains(t, err.Error(), "upstream request failed")
+	require.Equal(t, http.StatusBadGateway, rec.Code)
+	rawBody, ok := c.Get(OpsUpstreamRequestBodyKey)
+	require.True(t, ok)
+	_, ok = rawBody.([]byte)
+	require.True(t, ok)
+}
+
+func TestGatewayService_AnthropicAPIKeyPassthrough_ForwardDirect_EmptyResponseBody(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages", nil)
+
+	upstream := &anthropicHTTPUpstreamRecorder{
+		resp: &http.Response{
+			StatusCode: http.StatusOK,
+			Header:     http.Header{"x-request-id": []string{"rid-empty-body"}},
+			Body:       nil,
+		},
+	}
+	svc := &GatewayService{
+		cfg: &config.Config{
+			Security: config.SecurityConfig{
+				URLAllowlist: config.URLAllowlistConfig{Enabled: false},
+			},
+		},
+		httpUpstream: upstream,
+	}
+
+	result, err := svc.forwardAnthropicAPIKeyPassthrough(context.Background(), c, newAnthropicAPIKeyAccountForTest(), []byte(`{"model":"x"}`), "x", false, time.Now())
+	require.Nil(t, result)
+	require.Error(t, err)
+	require.Contains(t, err.Error(), "empty response")
+}
+
+func TestExtractAnthropicSSEDataLine(t *testing.T) {
+	t.Run("valid data line with spaces", func(t *testing.T) {
+		data, ok := extractAnthropicSSEDataLine("data:   {\"type\":\"message_start\"}")
+		require.True(t, ok)
+		require.Equal(t, `{"type":"message_start"}`, data)
+	})
+
+	t.Run("non data line", func(t *testing.T) {
+		data, ok := extractAnthropicSSEDataLine("event: message_start")
+		require.False(t, ok)
+		require.Empty(t, data)
+	})
+}
+
+func TestGatewayService_ParseSSEUsagePassthrough_MessageStartFallbacks(t *testing.T) {
+	svc := &GatewayService{}
+	usage := &ClaudeUsage{}
+	data := `{"type":"message_start","message":{"usage":{"input_tokens":12,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"cached_tokens":9,"cache_creation":{"ephemeral_5m_input_tokens":3,"ephemeral_1h_input_tokens":4}}}}`
+
+	svc.parseSSEUsagePassthrough(data, usage)
+
+	require.Equal(t, 12, usage.InputTokens)
+	require.Equal(t, 9, usage.CacheReadInputTokens, "应兼容 cached_tokens 字段")
+	require.Equal(t, 7, usage.CacheCreationInputTokens, "聚合字段为空时应从 5m/1h 明细回填")
+	require.Equal(t, 3, usage.CacheCreation5mTokens)
+	require.Equal(t, 4, usage.CacheCreation1hTokens)
+}
+
+func TestGatewayService_ParseSSEUsagePassthrough_MessageDeltaSelectiveOverwrite(t *testing.T) {
+	svc := &GatewayService{}
+	usage := &ClaudeUsage{
+		InputTokens:           10,
+		CacheCreation5mTokens: 2,
+		CacheCreation1hTokens: 6,
+	}
+	data := `{"type":"message_delta","usage":{"input_tokens":0,"output_tokens":5,"cache_creation_input_tokens":8,"cache_read_input_tokens":0,"cached_tokens":11,"cache_creation":{"ephemeral_5m_input_tokens":1,"ephemeral_1h_input_tokens":0}}}`
+
+	svc.parseSSEUsagePassthrough(data, usage)
+
+	require.Equal(t, 10, usage.InputTokens, "message_delta 中 0 值不应覆盖已有 input_tokens")
+	require.Equal(t, 5, usage.OutputTokens)
+	require.Equal(t, 8, usage.CacheCreationInputTokens)
+	require.Equal(t, 11, usage.CacheReadInputTokens, "cache_read_input_tokens 为空时应回退到 cached_tokens")
+	require.Equal(t, 1, usage.CacheCreation5mTokens)
+	require.Equal(t, 6, usage.CacheCreation1hTokens, "message_delta 中 0 值不应覆盖已有 1h 明细")
+}
+
+func TestGatewayService_ParseSSEUsagePassthrough_NoopCases(t *testing.T) {
+	svc := &GatewayService{}
+
+	usage := &ClaudeUsage{InputTokens: 3}
+	svc.parseSSEUsagePassthrough("", usage)
+	require.Equal(t, 3, usage.InputTokens)
+
+	svc.parseSSEUsagePassthrough("[DONE]", usage)
+	require.Equal(t, 3, usage.InputTokens)
+
+	svc.parseSSEUsagePassthrough("not-json", usage)
+	require.Equal(t, 3, usage.InputTokens)
+
+	// nil usage 不应 panic
+	svc.parseSSEUsagePassthrough(`{"type":"message_start"}`, nil)
+}
+
+func TestGatewayService_ParseSSEUsagePassthrough_FallbackFromUsageNode(t *testing.T) {
+	svc := &GatewayService{}
+	usage := &ClaudeUsage{}
+	data := `{"type":"content_block_delta","usage":{"cached_tokens":6,"cache_creation":{"ephemeral_5m_input_tokens":2,"ephemeral_1h_input_tokens":1}}}`
+
+	svc.parseSSEUsagePassthrough(data, usage)
+
+	require.Equal(t, 6, usage.CacheReadInputTokens)
+	require.Equal(t, 3, usage.CacheCreationInputTokens)
+}
+
+func TestParseClaudeUsageFromResponseBody(t *testing.T) {
+	t.Run("empty or missing usage", func(t *testing.T) {
+		got := parseClaudeUsageFromResponseBody(nil)
+		require.NotNil(t, got)
+		require.Equal(t, 0, got.InputTokens)
+
+		got = parseClaudeUsageFromResponseBody([]byte(`{"id":"x"}`))
+		require.NotNil(t, got)
+		require.Equal(t, 0, got.OutputTokens)
+	})
+
+	t.Run("parse all usage fields and fallback", func(t *testing.T) {
+		body := []byte(`{"usage":{"input_tokens":21,"output_tokens":34,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"cached_tokens":13,"cache_creation":{"ephemeral_5m_input_tokens":5,"ephemeral_1h_input_tokens":8}}}`)
+		got := parseClaudeUsageFromResponseBody(body)
+		require.Equal(t, 21, got.InputTokens)
+		require.Equal(t, 34, got.OutputTokens)
+		require.Equal(t, 13, got.CacheReadInputTokens, "cache_read_input_tokens 为空时应回退 cached_tokens")
+		require.Equal(t, 13, got.CacheCreationInputTokens, "聚合字段为空时应由 5m/1h 回填")
+		require.Equal(t, 5, got.CacheCreation5mTokens)
+		require.Equal(t, 8, got.CacheCreation1hTokens)
+	})
+
+	t.Run("keep explicit aggregate values", func(t *testing.T) {
+		body := []byte(`{"usage":{"input_tokens":1,"output_tokens":2,"cache_creation_input_tokens":9,"cache_read_input_tokens":7,"cached_tokens":99,"cache_creation":{"ephemeral_5m_input_tokens":4,"ephemeral_1h_input_tokens":5}}}`)
+		got := parseClaudeUsageFromResponseBody(body)
+		require.Equal(t, 9, got.CacheCreationInputTokens, "已显式提供聚合字段时不应被明细覆盖")
+		require.Equal(t, 7, got.CacheReadInputTokens, "已显式提供 cache_read_input_tokens 时不应回退 cached_tokens")
+	})
+}
+
+func TestGatewayService_AnthropicAPIKeyPassthrough_StreamingErrTooLong(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages", nil)
+
+	svc := &GatewayService{
+		cfg: &config.Config{
+			Gateway: config.GatewayConfig{
+				MaxLineSize: 32,
+			},
+		},
+	}
+
+	// Scanner 初始缓冲为 64KB，构造更长单行触发 bufio.ErrTooLong。
+	longLine := "data: " + strings.Repeat("x", 80*1024)
+	resp := &http.Response{
+		StatusCode: http.StatusOK,
+		Header:     http.Header{"Content-Type": []string{"text/event-stream"}},
+		Body:       io.NopCloser(strings.NewReader(longLine)),
+	}
+
+	result, err := svc.handleStreamingResponseAnthropicAPIKeyPassthrough(context.Background(), resp, c, &Account{ID: 2}, time.Now(), "claude-3-7-sonnet-20250219")
+	require.Error(t, err)
+	require.ErrorIs(t, err, bufio.ErrTooLong)
+	require.NotNil(t, result)
+}
+
+func TestGatewayService_AnthropicAPIKeyPassthrough_StreamingDataIntervalTimeout(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages", nil)
+
+	svc := &GatewayService{
+		cfg: &config.Config{
+			Gateway: config.GatewayConfig{
+				StreamDataIntervalTimeout: 1,
+				MaxLineSize:               defaultMaxLineSize,
+			},
+		},
+		rateLimitService: &RateLimitService{},
+	}
+
+	pr, pw := io.Pipe()
+	resp := &http.Response{
+		StatusCode: http.StatusOK,
+		Header:     http.Header{"Content-Type": []string{"text/event-stream"}},
+		Body:       pr,
+	}
+
+	result, err := svc.handleStreamingResponseAnthropicAPIKeyPassthrough(context.Background(), resp, c, &Account{ID: 5}, time.Now(), "claude-3-7-sonnet-20250219")
+	_ = pw.Close()
+	_ = pr.Close()
+
+	require.Error(t, err)
+	require.Contains(t, err.Error(), "stream data interval timeout")
+	require.NotNil(t, result)
+	require.False(t, result.clientDisconnect)
+}
+
+func TestGatewayService_AnthropicAPIKeyPassthrough_StreamingReadError(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages", nil)
+
+	svc := &GatewayService{
+		cfg: &config.Config{
+			Gateway: config.GatewayConfig{
+				MaxLineSize: defaultMaxLineSize,
+			},
+		},
+	}
+
+	resp := &http.Response{
+		StatusCode: http.StatusOK,
+		Header:     http.Header{"Content-Type": []string{"text/event-stream"}},
+		Body: &streamReadCloser{
+			err: io.ErrUnexpectedEOF,
+		},
+	}
+
+	result, err := svc.handleStreamingResponseAnthropicAPIKeyPassthrough(context.Background(), resp, c, &Account{ID: 6}, time.Now(), "claude-3-7-sonnet-20250219")
+	require.Error(t, err)
+	require.Contains(t, err.Error(), "stream read error")
+	require.NotNil(t, result)
+	require.False(t, result.clientDisconnect)
+}
+
+func TestGatewayService_AnthropicAPIKeyPassthrough_StreamingTimeoutAfterClientDisconnect(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages", nil)
+	c.Writer = &failWriteResponseWriter{ResponseWriter: c.Writer}
+
+	svc := &GatewayService{
+		cfg: &config.Config{
+			Gateway: config.GatewayConfig{
+				StreamDataIntervalTimeout: 1,
+				MaxLineSize:               defaultMaxLineSize,
+			},
+		},
+		rateLimitService: &RateLimitService{},
+	}
+
+	pr, pw := io.Pipe()
+	resp := &http.Response{
+		StatusCode: http.StatusOK,
+		Header:     http.Header{"Content-Type": []string{"text/event-stream"}},
+		Body:       pr,
+	}
+
+	done := make(chan struct{})
+	go func() {
+		defer close(done)
+		_, _ = pw.Write([]byte(`data: {"type":"message_start","message":{"usage":{"input_tokens":9}}}` + "\n"))
+		// 保持上游连接静默，触发数据间隔超时分支。
+		time.Sleep(1500 * time.Millisecond)
+		_ = pw.Close()
+	}()
+
+	result, err := svc.handleStreamingResponseAnthropicAPIKeyPassthrough(context.Background(), resp, c, &Account{ID: 7}, time.Now(), "claude-3-7-sonnet-20250219")
+	_ = pr.Close()
+	<-done
+
+	require.NoError(t, err)
+	require.NotNil(t, result)
+	require.True(t, result.clientDisconnect)
+	require.Equal(t, 9, result.usage.InputTokens)
+}
+
+func TestGatewayService_AnthropicAPIKeyPassthrough_StreamingContextCanceled(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages", nil)
+
+	svc := &GatewayService{
+		cfg: &config.Config{
+			Gateway: config.GatewayConfig{
+				MaxLineSize: defaultMaxLineSize,
+			},
+		},
+	}
+
+	resp := &http.Response{
+		StatusCode: http.StatusOK,
+		Header:     http.Header{"Content-Type": []string{"text/event-stream"}},
+		Body: &streamReadCloser{
+			err: context.Canceled,
+		},
+	}
+
+	result, err := svc.handleStreamingResponseAnthropicAPIKeyPassthrough(context.Background(), resp, c, &Account{ID: 3}, time.Now(), "claude-3-7-sonnet-20250219")
+	require.NoError(t, err)
+	require.NotNil(t, result)
+	require.True(t, result.clientDisconnect)
+}
+
+func TestGatewayService_AnthropicAPIKeyPassthrough_StreamingUpstreamReadErrorAfterClientDisconnect(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages", nil)
+	c.Writer = &failWriteResponseWriter{ResponseWriter: c.Writer}
+
+	svc := &GatewayService{
+		cfg: &config.Config{
+			Gateway: config.GatewayConfig{
+				MaxLineSize: defaultMaxLineSize,
+			},
+		},
+	}
+
+	resp := &http.Response{
+		StatusCode: http.StatusOK,
+		Header:     http.Header{"Content-Type": []string{"text/event-stream"}},
+		Body: &streamReadCloser{
+			payload: []byte(`data: {"type":"message_start","message":{"usage":{"input_tokens":8}}}` + "\n\n"),
+			err:     io.ErrUnexpectedEOF,
+		},
+	}
+
+	result, err := svc.handleStreamingResponseAnthropicAPIKeyPassthrough(context.Background(), resp, c, &Account{ID: 4}, time.Now(), "claude-3-7-sonnet-20250219")
+	require.NoError(t, err)
+	require.NotNil(t, result)
+	require.True(t, result.clientDisconnect)
+	require.Equal(t, 8, result.usage.InputTokens)
+}

backend/internal/service/gateway_service.go

@@ -3041,6 +3041,10 @@ func (s *GatewayService) Forward(ctx context.Context, c *gin.Context, account *A
 		return nil, fmt.Errorf("parse request: empty request")
 	}
 
+	if account != nil && account.IsAnthropicAPIKeyPassthroughEnabled() {
+		return s.forwardAnthropicAPIKeyPassthrough(ctx, c, account, parsed.Body, parsed.Model, parsed.Stream, startTime)
+	}
+
 	body := parsed.Body
 	reqModel := parsed.Model
 	reqStream := parsed.Stream
@@ -3120,14 +3124,14 @@ func (s *GatewayService) Forward(ctx context.Context, c *gin.Context, account *A
 	// 调试日志：记录即将转发的账号信息
 	logger.LegacyPrintf("service.gateway", "[Forward] Using account: ID=%d Name=%s Platform=%s Type=%s TLSFingerprint=%v Proxy=%s",
 		account.ID, account.Name, account.Platform, account.Type, account.IsTLSFingerprintEnabled(), proxyURL)
+	// 重试间复用同一请求体，避免每次 string(body) 产生额外分配。
+	setOpsUpstreamRequestBody(c, body)
 
 	// 重试循环
 	var resp *http.Response
 	retryStart := time.Now()
 	for attempt := 1; attempt <= maxRetryAttempts; attempt++ {
 		// 构建上游请求（每次重试需要重新构建，因为请求体需要重新读取）
-		// Capture upstream request body for ops retry of this attempt.
-		c.Set(OpsUpstreamRequestBodyKey, string(body))
 		upstreamReq, err := s.buildUpstreamRequest(ctx, c, account, body, token, tokenType, reqModel, reqStream, shouldMimicClaudeCode)
 		if err != nil {
 			return nil, err
@@ -3377,118 +3381,714 @@ func (s *GatewayService) Forward(ctx context.Context, c *gin.Context, account *A
 			})
 			return nil, &UpstreamFailoverError{StatusCode: resp.StatusCode, ResponseBody: respBody}
 		}
-		return s.handleRetryExhaustedError(ctx, resp, c, account)
+		return s.handleRetryExhaustedError(ctx, resp, c, account)
+	}
+
+	// 处理可切换账号的错误
+	if resp.StatusCode >= 400 && s.shouldFailoverUpstreamError(resp.StatusCode) {
+		respBody, _ := io.ReadAll(io.LimitReader(resp.Body, 2<<20))
+		_ = resp.Body.Close()
+		resp.Body = io.NopCloser(bytes.NewReader(respBody))
+
+		// 调试日志：打印上游错误响应
+		logger.LegacyPrintf("service.gateway", "[Forward] Upstream error (failover): Account=%d(%s) Status=%d RequestID=%s Body=%s",
+			account.ID, account.Name, resp.StatusCode, resp.Header.Get("x-request-id"), truncateString(string(respBody), 1000))
+
+		s.handleFailoverSideEffects(ctx, resp, account)
+		appendOpsUpstreamError(c, OpsUpstreamErrorEvent{
+			Platform:           account.Platform,
+			AccountID:          account.ID,
+			UpstreamStatusCode: resp.StatusCode,
+			UpstreamRequestID:  resp.Header.Get("x-request-id"),
+			Kind:               "failover",
+			Message:            extractUpstreamErrorMessage(respBody),
+			Detail: func() string {
+				if s.cfg != nil && s.cfg.Gateway.LogUpstreamErrorBody {
+					return truncateString(string(respBody), s.cfg.Gateway.LogUpstreamErrorBodyMaxBytes)
+				}
+				return ""
+			}(),
+		})
+		return nil, &UpstreamFailoverError{StatusCode: resp.StatusCode, ResponseBody: respBody}
+	}
+	if resp.StatusCode >= 400 {
+		// 可选：对部分 400 触发 failover（默认关闭以保持语义）
+		if resp.StatusCode == 400 && s.cfg != nil && s.cfg.Gateway.FailoverOn400 {
+			respBody, readErr := io.ReadAll(io.LimitReader(resp.Body, 2<<20))
+			if readErr != nil {
+				// ReadAll failed, fall back to normal error handling without consuming the stream
+				return s.handleErrorResponse(ctx, resp, c, account)
+			}
+			_ = resp.Body.Close()
+			resp.Body = io.NopCloser(bytes.NewReader(respBody))
+
+			if s.shouldFailoverOn400(respBody) {
+				upstreamMsg := strings.TrimSpace(extractUpstreamErrorMessage(respBody))
+				upstreamMsg = sanitizeUpstreamErrorMessage(upstreamMsg)
+				upstreamDetail := ""
+				if s.cfg != nil && s.cfg.Gateway.LogUpstreamErrorBody {
+					maxBytes := s.cfg.Gateway.LogUpstreamErrorBodyMaxBytes
+					if maxBytes <= 0 {
+						maxBytes = 2048
+					}
+					upstreamDetail = truncateString(string(respBody), maxBytes)
+				}
+				appendOpsUpstreamError(c, OpsUpstreamErrorEvent{
+					Platform:           account.Platform,
+					AccountID:          account.ID,
+					AccountName:        account.Name,
+					UpstreamStatusCode: resp.StatusCode,
+					UpstreamRequestID:  resp.Header.Get("x-request-id"),
+					Kind:               "failover_on_400",
+					Message:            upstreamMsg,
+					Detail:             upstreamDetail,
+				})
+
+				if s.cfg.Gateway.LogUpstreamErrorBody {
+					logger.LegacyPrintf("service.gateway",
+						"Account %d: 400 error, attempting failover: %s",
+						account.ID,
+						truncateForLog(respBody, s.cfg.Gateway.LogUpstreamErrorBodyMaxBytes),
+					)
+				} else {
+					logger.LegacyPrintf("service.gateway", "Account %d: 400 error, attempting failover", account.ID)
+				}
+				s.handleFailoverSideEffects(ctx, resp, account)
+				return nil, &UpstreamFailoverError{StatusCode: resp.StatusCode, ResponseBody: respBody}
+			}
+		}
+		return s.handleErrorResponse(ctx, resp, c, account)
+	}
+
+	// 处理正常响应
+	var usage *ClaudeUsage
+	var firstTokenMs *int
+	var clientDisconnect bool
+	if reqStream {
+		streamResult, err := s.handleStreamingResponse(ctx, resp, c, account, startTime, originalModel, reqModel, shouldMimicClaudeCode)
+		if err != nil {
+			if err.Error() == "have error in stream" {
+				return nil, &UpstreamFailoverError{
+					StatusCode: 403,
+				}
+			}
+			return nil, err
+		}
+		usage = streamResult.usage
+		firstTokenMs = streamResult.firstTokenMs
+		clientDisconnect = streamResult.clientDisconnect
+	} else {
+		usage, err = s.handleNonStreamingResponse(ctx, resp, c, account, originalModel, reqModel)
+		if err != nil {
+			return nil, err
+		}
+	}
+
+	return &ForwardResult{
+		RequestID:        resp.Header.Get("x-request-id"),
+		Usage:            *usage,
+		Model:            originalModel, // 使用原始模型用于计费和日志
+		Stream:           reqStream,
+		Duration:         time.Since(startTime),
+		FirstTokenMs:     firstTokenMs,
+		ClientDisconnect: clientDisconnect,
+	}, nil
+}
+
+func (s *GatewayService) forwardAnthropicAPIKeyPassthrough(
+	ctx context.Context,
+	c *gin.Context,
+	account *Account,
+	body []byte,
+	reqModel string,
+	reqStream bool,
+	startTime time.Time,
+) (*ForwardResult, error) {
+	token, tokenType, err := s.GetAccessToken(ctx, account)
+	if err != nil {
+		return nil, err
+	}
+	if tokenType != "apikey" {
+		return nil, fmt.Errorf("anthropic api key passthrough requires apikey token, got: %s", tokenType)
+	}
+
+	proxyURL := ""
+	if account.ProxyID != nil && account.Proxy != nil {
+		proxyURL = account.Proxy.URL()
+	}
+
+	logger.LegacyPrintf("service.gateway", "[Anthropic 自动透传] 命中 API Key 透传分支: account=%d name=%s model=%s stream=%v",
+		account.ID, account.Name, reqModel, reqStream)
+
+	if c != nil {
+		c.Set("anthropic_passthrough", true)
+	}
+	// 重试间复用同一请求体，避免每次 string(body) 产生额外分配。
+	setOpsUpstreamRequestBody(c, body)
+
+	var resp *http.Response
+	retryStart := time.Now()
+	for attempt := 1; attempt <= maxRetryAttempts; attempt++ {
+		upstreamReq, err := s.buildUpstreamRequestAnthropicAPIKeyPassthrough(ctx, c, account, body, token)
+		if err != nil {
+			return nil, err
+		}
+
+		resp, err = s.httpUpstream.DoWithTLS(upstreamReq, proxyURL, account.ID, account.Concurrency, account.IsTLSFingerprintEnabled())
+		if err != nil {
+			if resp != nil && resp.Body != nil {
+				_ = resp.Body.Close()
+			}
+			safeErr := sanitizeUpstreamErrorMessage(err.Error())
+			setOpsUpstreamError(c, 0, safeErr, "")
+			appendOpsUpstreamError(c, OpsUpstreamErrorEvent{
+				Platform:           account.Platform,
+				AccountID:          account.ID,
+				AccountName:        account.Name,
+				UpstreamStatusCode: 0,
+				Passthrough:        true,
+				Kind:               "request_error",
+				Message:            safeErr,
+			})
+			c.JSON(http.StatusBadGateway, gin.H{
+				"type": "error",
+				"error": gin.H{
+					"type":    "upstream_error",
+					"message": "Upstream request failed",
+				},
+			})
+			return nil, fmt.Errorf("upstream request failed: %s", safeErr)
+		}
+
+		// 透传分支禁止 400 请求体降级重试（该重试会改写请求体）
+		if resp.StatusCode >= 400 && resp.StatusCode != 400 && s.shouldRetryUpstreamError(account, resp.StatusCode) {
+			if attempt < maxRetryAttempts {
+				elapsed := time.Since(retryStart)
+				if elapsed >= maxRetryElapsed {
+					break
+				}
+
+				delay := retryBackoffDelay(attempt)
+				remaining := maxRetryElapsed - elapsed
+				if delay > remaining {
+					delay = remaining
+				}
+				if delay <= 0 {
+					break
+				}
+
+				respBody, _ := io.ReadAll(io.LimitReader(resp.Body, 2<<20))
+				_ = resp.Body.Close()
+				appendOpsUpstreamError(c, OpsUpstreamErrorEvent{
+					Platform:           account.Platform,
+					AccountID:          account.ID,
+					AccountName:        account.Name,
+					UpstreamStatusCode: resp.StatusCode,
+					UpstreamRequestID:  resp.Header.Get("x-request-id"),
+					Passthrough:        true,
+					Kind:               "retry",
+					Message:            extractUpstreamErrorMessage(respBody),
+					Detail: func() string {
+						if s.cfg != nil && s.cfg.Gateway.LogUpstreamErrorBody {
+							return truncateString(string(respBody), s.cfg.Gateway.LogUpstreamErrorBodyMaxBytes)
+						}
+						return ""
+					}(),
+				})
+				logger.LegacyPrintf("service.gateway", "Anthropic passthrough account %d: upstream error %d, retry %d/%d after %v (elapsed=%v/%v)",
+					account.ID, resp.StatusCode, attempt, maxRetryAttempts, delay, elapsed, maxRetryElapsed)
+				if err := sleepWithContext(ctx, delay); err != nil {
+					return nil, err
+				}
+				continue
+			}
+			break
+		}
+
+		break
+	}
+	if resp == nil || resp.Body == nil {
+		return nil, errors.New("upstream request failed: empty response")
+	}
+	defer func() { _ = resp.Body.Close() }()
+
+	if resp.StatusCode >= 400 && s.shouldRetryUpstreamError(account, resp.StatusCode) {
+		if s.shouldFailoverUpstreamError(resp.StatusCode) {
+			respBody, _ := io.ReadAll(io.LimitReader(resp.Body, 2<<20))
+			_ = resp.Body.Close()
+			resp.Body = io.NopCloser(bytes.NewReader(respBody))
+
+			logger.LegacyPrintf("service.gateway", "[Anthropic Passthrough] Upstream error (retry exhausted, failover): Account=%d(%s) Status=%d RequestID=%s Body=%s",
+				account.ID, account.Name, resp.StatusCode, resp.Header.Get("x-request-id"), truncateString(string(respBody), 1000))
+
+			s.handleRetryExhaustedSideEffects(ctx, resp, account)
+			appendOpsUpstreamError(c, OpsUpstreamErrorEvent{
+				Platform:           account.Platform,
+				AccountID:          account.ID,
+				AccountName:        account.Name,
+				UpstreamStatusCode: resp.StatusCode,
+				UpstreamRequestID:  resp.Header.Get("x-request-id"),
+				Passthrough:        true,
+				Kind:               "retry_exhausted_failover",
+				Message:            extractUpstreamErrorMessage(respBody),
+				Detail: func() string {
+					if s.cfg != nil && s.cfg.Gateway.LogUpstreamErrorBody {
+						return truncateString(string(respBody), s.cfg.Gateway.LogUpstreamErrorBodyMaxBytes)
+					}
+					return ""
+				}(),
+			})
+			return nil, &UpstreamFailoverError{StatusCode: resp.StatusCode, ResponseBody: respBody}
+		}
+		return s.handleRetryExhaustedError(ctx, resp, c, account)
+	}
+
+	if resp.StatusCode >= 400 && s.shouldFailoverUpstreamError(resp.StatusCode) {
+		respBody, _ := io.ReadAll(io.LimitReader(resp.Body, 2<<20))
+		_ = resp.Body.Close()
+		resp.Body = io.NopCloser(bytes.NewReader(respBody))
+
+		logger.LegacyPrintf("service.gateway", "[Anthropic Passthrough] Upstream error (failover): Account=%d(%s) Status=%d RequestID=%s Body=%s",
+			account.ID, account.Name, resp.StatusCode, resp.Header.Get("x-request-id"), truncateString(string(respBody), 1000))
+
+		s.handleFailoverSideEffects(ctx, resp, account)
+		appendOpsUpstreamError(c, OpsUpstreamErrorEvent{
+			Platform:           account.Platform,
+			AccountID:          account.ID,
+			AccountName:        account.Name,
+			UpstreamStatusCode: resp.StatusCode,
+			UpstreamRequestID:  resp.Header.Get("x-request-id"),
+			Passthrough:        true,
+			Kind:               "failover",
+			Message:            extractUpstreamErrorMessage(respBody),
+			Detail: func() string {
+				if s.cfg != nil && s.cfg.Gateway.LogUpstreamErrorBody {
+					return truncateString(string(respBody), s.cfg.Gateway.LogUpstreamErrorBodyMaxBytes)
+				}
+				return ""
+			}(),
+		})
+		return nil, &UpstreamFailoverError{StatusCode: resp.StatusCode, ResponseBody: respBody}
+	}
+
+	if resp.StatusCode >= 400 {
+		return s.handleErrorResponse(ctx, resp, c, account)
+	}
+
+	var usage *ClaudeUsage
+	var firstTokenMs *int
+	var clientDisconnect bool
+	if reqStream {
+		streamResult, err := s.handleStreamingResponseAnthropicAPIKeyPassthrough(ctx, resp, c, account, startTime, reqModel)
+		if err != nil {
+			return nil, err
+		}
+		usage = streamResult.usage
+		firstTokenMs = streamResult.firstTokenMs
+		clientDisconnect = streamResult.clientDisconnect
+	} else {
+		usage, err = s.handleNonStreamingResponseAnthropicAPIKeyPassthrough(ctx, resp, c, account)
+		if err != nil {
+			return nil, err
+		}
+	}
+	if usage == nil {
+		usage = &ClaudeUsage{}
+	}
+
+	return &ForwardResult{
+		RequestID:        resp.Header.Get("x-request-id"),
+		Usage:            *usage,
+		Model:            reqModel,
+		Stream:           reqStream,
+		Duration:         time.Since(startTime),
+		FirstTokenMs:     firstTokenMs,
+		ClientDisconnect: clientDisconnect,
+	}, nil
+}
+
+func (s *GatewayService) buildUpstreamRequestAnthropicAPIKeyPassthrough(
+	ctx context.Context,
+	c *gin.Context,
+	account *Account,
+	body []byte,
+	token string,
+) (*http.Request, error) {
+	targetURL := claudeAPIURL
+	baseURL := account.GetBaseURL()
+	if baseURL != "" {
+		validatedURL, err := s.validateUpstreamBaseURL(baseURL)
+		if err != nil {
+			return nil, err
+		}
+		targetURL = validatedURL + "/v1/messages"
+	}
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, targetURL, bytes.NewReader(body))
+	if err != nil {
+		return nil, err
+	}
+
+	if c != nil && c.Request != nil {
+		for key, values := range c.Request.Header {
+			lowerKey := strings.ToLower(strings.TrimSpace(key))
+			if !allowedHeaders[lowerKey] {
+				continue
+			}
+			for _, v := range values {
+				req.Header.Add(key, v)
+			}
+		}
+	}
+
+	// 覆盖入站鉴权残留，并注入上游认证
+	req.Header.Del("authorization")
+	req.Header.Del("x-api-key")
+	req.Header.Del("x-goog-api-key")
+	req.Header.Del("cookie")
+	req.Header.Set("x-api-key", token)
+
+	if req.Header.Get("content-type") == "" {
+		req.Header.Set("content-type", "application/json")
+	}
+	if req.Header.Get("anthropic-version") == "" {
+		req.Header.Set("anthropic-version", "2023-06-01")
+	}
+
+	return req, nil
+}
+
+func (s *GatewayService) handleStreamingResponseAnthropicAPIKeyPassthrough(
+	ctx context.Context,
+	resp *http.Response,
+	c *gin.Context,
+	account *Account,
+	startTime time.Time,
+	model string,
+) (*streamingResult, error) {
+	if s.rateLimitService != nil {
+		s.rateLimitService.UpdateSessionWindow(ctx, account, resp.Header)
+	}
+
+	writeAnthropicPassthroughResponseHeaders(c.Writer.Header(), resp.Header, s.cfg)
+
+	contentType := strings.TrimSpace(resp.Header.Get("Content-Type"))
+	if contentType == "" {
+		contentType = "text/event-stream"
+	}
+	c.Header("Content-Type", contentType)
+	if c.Writer.Header().Get("Cache-Control") == "" {
+		c.Header("Cache-Control", "no-cache")
+	}
+	if c.Writer.Header().Get("Connection") == "" {
+		c.Header("Connection", "keep-alive")
+	}
+	c.Header("X-Accel-Buffering", "no")
+	if v := resp.Header.Get("x-request-id"); v != "" {
+		c.Header("x-request-id", v)
+	}
+
+	w := c.Writer
+	flusher, ok := w.(http.Flusher)
+	if !ok {
+		return nil, errors.New("streaming not supported")
+	}
+
+	usage := &ClaudeUsage{}
+	var firstTokenMs *int
+	clientDisconnected := false
+
+	scanner := bufio.NewScanner(resp.Body)
+	maxLineSize := defaultMaxLineSize
+	if s.cfg != nil && s.cfg.Gateway.MaxLineSize > 0 {
+		maxLineSize = s.cfg.Gateway.MaxLineSize
+	}
+	scanBuf := getSSEScannerBuf64K()
+	scanner.Buffer(scanBuf[:0], maxLineSize)
+
+	type scanEvent struct {
+		line string
+		err  error
+	}
+	events := make(chan scanEvent, 16)
+	done := make(chan struct{})
+	sendEvent := func(ev scanEvent) bool {
+		select {
+		case events <- ev:
+			return true
+		case <-done:
+			return false
+		}
+	}
+	var lastReadAt int64
+	atomic.StoreInt64(&lastReadAt, time.Now().UnixNano())
+	go func(scanBuf *sseScannerBuf64K) {
+		defer putSSEScannerBuf64K(scanBuf)
+		defer close(events)
+		for scanner.Scan() {
+			atomic.StoreInt64(&lastReadAt, time.Now().UnixNano())
+			if !sendEvent(scanEvent{line: scanner.Text()}) {
+				return
+			}
+		}
+		if err := scanner.Err(); err != nil {
+			_ = sendEvent(scanEvent{err: err})
+		}
+	}(scanBuf)
+	defer close(done)
+
+	streamInterval := time.Duration(0)
+	if s.cfg != nil && s.cfg.Gateway.StreamDataIntervalTimeout > 0 {
+		streamInterval = time.Duration(s.cfg.Gateway.StreamDataIntervalTimeout) * time.Second
+	}
+	var intervalTicker *time.Ticker
+	if streamInterval > 0 {
+		intervalTicker = time.NewTicker(streamInterval)
+		defer intervalTicker.Stop()
+	}
+	var intervalCh <-chan time.Time
+	if intervalTicker != nil {
+		intervalCh = intervalTicker.C
+	}
+
+	for {
+		select {
+		case ev, ok := <-events:
+			if !ok {
+				if !clientDisconnected {
+					// 兜底补刷，确保最后一个未以空行结尾的事件也能及时送达客户端。
+					flusher.Flush()
+				}
+				return &streamingResult{usage: usage, firstTokenMs: firstTokenMs, clientDisconnect: clientDisconnected}, nil
+			}
+			if ev.err != nil {
+				if clientDisconnected {
+					logger.LegacyPrintf("service.gateway", "[Anthropic passthrough] Upstream read error after client disconnect: account=%d err=%v", account.ID, ev.err)
+					return &streamingResult{usage: usage, firstTokenMs: firstTokenMs, clientDisconnect: true}, nil
+				}
+				if errors.Is(ev.err, context.Canceled) || errors.Is(ev.err, context.DeadlineExceeded) {
+					logger.LegacyPrintf("service.gateway", "[Anthropic passthrough] 流读取被取消: account=%d request_id=%s err=%v ctx_err=%v",
+						account.ID, resp.Header.Get("x-request-id"), ev.err, ctx.Err())
+					return &streamingResult{usage: usage, firstTokenMs: firstTokenMs, clientDisconnect: true}, nil
+				}
+				if errors.Is(ev.err, bufio.ErrTooLong) {
+					logger.LegacyPrintf("service.gateway", "[Anthropic passthrough] SSE line too long: account=%d max_size=%d error=%v", account.ID, maxLineSize, ev.err)
+					return &streamingResult{usage: usage, firstTokenMs: firstTokenMs}, ev.err
+				}
+				return &streamingResult{usage: usage, firstTokenMs: firstTokenMs}, fmt.Errorf("stream read error: %w", ev.err)
+			}
+
+			line := ev.line
+			if data, ok := extractAnthropicSSEDataLine(line); ok {
+				trimmed := strings.TrimSpace(data)
+				if firstTokenMs == nil && trimmed != "" && trimmed != "[DONE]" {
+					ms := int(time.Since(startTime).Milliseconds())
+					firstTokenMs = &ms
+				}
+				s.parseSSEUsagePassthrough(data, usage)
+			}
+
+			if !clientDisconnected {
+				if _, err := io.WriteString(w, line); err != nil {
+					clientDisconnected = true
+					logger.LegacyPrintf("service.gateway", "[Anthropic passthrough] Client disconnected during streaming, continue draining upstream for usage: account=%d", account.ID)
+				} else if _, err := io.WriteString(w, "\n"); err != nil {
+					clientDisconnected = true
+					logger.LegacyPrintf("service.gateway", "[Anthropic passthrough] Client disconnected during streaming, continue draining upstream for usage: account=%d", account.ID)
+				} else if line == "" {
+					// 按 SSE 事件边界刷出，减少每行 flush 带来的 syscall 开销。
+					flusher.Flush()
+				}
+			}
+
+		case <-intervalCh:
+			lastRead := time.Unix(0, atomic.LoadInt64(&lastReadAt))
+			if time.Since(lastRead) < streamInterval {
+				continue
+			}
+			if clientDisconnected {
+				logger.LegacyPrintf("service.gateway", "[Anthropic passthrough] Upstream timeout after client disconnect: account=%d model=%s", account.ID, model)
+				return &streamingResult{usage: usage, firstTokenMs: firstTokenMs, clientDisconnect: true}, nil
+			}
+			logger.LegacyPrintf("service.gateway", "[Anthropic passthrough] Stream data interval timeout: account=%d model=%s interval=%s", account.ID, model, streamInterval)
+			if s.rateLimitService != nil {
+				s.rateLimitService.HandleStreamTimeout(ctx, account, model)
+			}
+			return &streamingResult{usage: usage, firstTokenMs: firstTokenMs}, fmt.Errorf("stream data interval timeout")
+		}
+	}
+}
+
+func extractAnthropicSSEDataLine(line string) (string, bool) {
+	if !strings.HasPrefix(line, "data:") {
+		return "", false
+	}
+	start := len("data:")
+	for start < len(line) {
+		if line[start] != ' ' && line[start] != '\t' {
+			break
+		}
+		start++
+	}
+	return line[start:], true
+}
+
+func (s *GatewayService) parseSSEUsagePassthrough(data string, usage *ClaudeUsage) {
+	if usage == nil || data == "" || data == "[DONE]" {
+		return
+	}
+
+	parsed := gjson.Parse(data)
+	switch parsed.Get("type").String() {
+	case "message_start":
+		msgUsage := parsed.Get("message.usage")
+		if msgUsage.Exists() {
+			usage.InputTokens = int(msgUsage.Get("input_tokens").Int())
+			usage.CacheCreationInputTokens = int(msgUsage.Get("cache_creation_input_tokens").Int())
+			usage.CacheReadInputTokens = int(msgUsage.Get("cache_read_input_tokens").Int())
+
+			// 保持与通用解析一致：message_start 允许覆盖 5m/1h 明细（包括 0）。
+			cc5m := msgUsage.Get("cache_creation.ephemeral_5m_input_tokens")
+			cc1h := msgUsage.Get("cache_creation.ephemeral_1h_input_tokens")
+			if cc5m.Exists() || cc1h.Exists() {
+				usage.CacheCreation5mTokens = int(cc5m.Int())
+				usage.CacheCreation1hTokens = int(cc1h.Int())
+			}
+		}
+	case "message_delta":
+		deltaUsage := parsed.Get("usage")
+		if deltaUsage.Exists() {
+			if v := deltaUsage.Get("input_tokens").Int(); v > 0 {
+				usage.InputTokens = int(v)
+			}
+			if v := deltaUsage.Get("output_tokens").Int(); v > 0 {
+				usage.OutputTokens = int(v)
+			}
+			if v := deltaUsage.Get("cache_creation_input_tokens").Int(); v > 0 {
+				usage.CacheCreationInputTokens = int(v)
+			}
+			if v := deltaUsage.Get("cache_read_input_tokens").Int(); v > 0 {
+				usage.CacheReadInputTokens = int(v)
+			}
+
+			cc5m := deltaUsage.Get("cache_creation.ephemeral_5m_input_tokens")
+			cc1h := deltaUsage.Get("cache_creation.ephemeral_1h_input_tokens")
+			if cc5m.Exists() && cc5m.Int() > 0 {
+				usage.CacheCreation5mTokens = int(cc5m.Int())
+			}
+			if cc1h.Exists() && cc1h.Int() > 0 {
+				usage.CacheCreation1hTokens = int(cc1h.Int())
+			}
+		}
 	}
 
-	// 处理可切换账号的错误
-	if resp.StatusCode >= 400 && s.shouldFailoverUpstreamError(resp.StatusCode) {
-		respBody, _ := io.ReadAll(io.LimitReader(resp.Body, 2<<20))
-		_ = resp.Body.Close()
-		resp.Body = io.NopCloser(bytes.NewReader(respBody))
+	if usage.CacheReadInputTokens == 0 {
+		if cached := parsed.Get("message.usage.cached_tokens").Int(); cached > 0 {
+			usage.CacheReadInputTokens = int(cached)
+		}
+		if cached := parsed.Get("usage.cached_tokens").Int(); usage.CacheReadInputTokens == 0 && cached > 0 {
+			usage.CacheReadInputTokens = int(cached)
+		}
+	}
+	if usage.CacheCreationInputTokens == 0 {
+		cc5m := parsed.Get("message.usage.cache_creation.ephemeral_5m_input_tokens").Int()
+		cc1h := parsed.Get("message.usage.cache_creation.ephemeral_1h_input_tokens").Int()
+		if cc5m == 0 && cc1h == 0 {
+			cc5m = parsed.Get("usage.cache_creation.ephemeral_5m_input_tokens").Int()
+			cc1h = parsed.Get("usage.cache_creation.ephemeral_1h_input_tokens").Int()
+		}
+		total := cc5m + cc1h
+		if total > 0 {
+			usage.CacheCreationInputTokens = int(total)
+		}
+	}
+}
 
-		// 调试日志：打印上游错误响应
-		logger.LegacyPrintf("service.gateway", "[Forward] Upstream error (failover): Account=%d(%s) Status=%d RequestID=%s Body=%s",
-			account.ID, account.Name, resp.StatusCode, resp.Header.Get("x-request-id"), truncateString(string(respBody), 1000))
+func parseClaudeUsageFromResponseBody(body []byte) *ClaudeUsage {
+	usage := &ClaudeUsage{}
+	if len(body) == 0 {
+		return usage
+	}
 
-		s.handleFailoverSideEffects(ctx, resp, account)
-		appendOpsUpstreamError(c, OpsUpstreamErrorEvent{
-			Platform:           account.Platform,
-			AccountID:          account.ID,
-			UpstreamStatusCode: resp.StatusCode,
-			UpstreamRequestID:  resp.Header.Get("x-request-id"),
-			Kind:               "failover",
-			Message:            extractUpstreamErrorMessage(respBody),
-			Detail: func() string {
-				if s.cfg != nil && s.cfg.Gateway.LogUpstreamErrorBody {
-					return truncateString(string(respBody), s.cfg.Gateway.LogUpstreamErrorBodyMaxBytes)
-				}
-				return ""
-			}(),
-		})
-		return nil, &UpstreamFailoverError{StatusCode: resp.StatusCode, ResponseBody: respBody}
+	parsed := gjson.ParseBytes(body)
+	usageNode := parsed.Get("usage")
+	if !usageNode.Exists() {
+		return usage
 	}
-	if resp.StatusCode >= 400 {
-		// 可选：对部分 400 触发 failover（默认关闭以保持语义）
-		if resp.StatusCode == 400 && s.cfg != nil && s.cfg.Gateway.FailoverOn400 {
-			respBody, readErr := io.ReadAll(io.LimitReader(resp.Body, 2<<20))
-			if readErr != nil {
-				// ReadAll failed, fall back to normal error handling without consuming the stream
-				return s.handleErrorResponse(ctx, resp, c, account)
-			}
-			_ = resp.Body.Close()
-			resp.Body = io.NopCloser(bytes.NewReader(respBody))
 
-			if s.shouldFailoverOn400(respBody) {
-				upstreamMsg := strings.TrimSpace(extractUpstreamErrorMessage(respBody))
-				upstreamMsg = sanitizeUpstreamErrorMessage(upstreamMsg)
-				upstreamDetail := ""
-				if s.cfg != nil && s.cfg.Gateway.LogUpstreamErrorBody {
-					maxBytes := s.cfg.Gateway.LogUpstreamErrorBodyMaxBytes
-					if maxBytes <= 0 {
-						maxBytes = 2048
-					}
-					upstreamDetail = truncateString(string(respBody), maxBytes)
-				}
-				appendOpsUpstreamError(c, OpsUpstreamErrorEvent{
-					Platform:           account.Platform,
-					AccountID:          account.ID,
-					AccountName:        account.Name,
-					UpstreamStatusCode: resp.StatusCode,
-					UpstreamRequestID:  resp.Header.Get("x-request-id"),
-					Kind:               "failover_on_400",
-					Message:            upstreamMsg,
-					Detail:             upstreamDetail,
-				})
+	usage.InputTokens = int(usageNode.Get("input_tokens").Int())
+	usage.OutputTokens = int(usageNode.Get("output_tokens").Int())
+	usage.CacheCreationInputTokens = int(usageNode.Get("cache_creation_input_tokens").Int())
+	usage.CacheReadInputTokens = int(usageNode.Get("cache_read_input_tokens").Int())
 
-				if s.cfg.Gateway.LogUpstreamErrorBody {
-					logger.LegacyPrintf("service.gateway",
-						"Account %d: 400 error, attempting failover: %s",
-						account.ID,
-						truncateForLog(respBody, s.cfg.Gateway.LogUpstreamErrorBodyMaxBytes),
-					)
-				} else {
-					logger.LegacyPrintf("service.gateway", "Account %d: 400 error, attempting failover", account.ID)
-				}
-				s.handleFailoverSideEffects(ctx, resp, account)
-				return nil, &UpstreamFailoverError{StatusCode: resp.StatusCode, ResponseBody: respBody}
-			}
+	cc5m := usageNode.Get("cache_creation.ephemeral_5m_input_tokens").Int()
+	cc1h := usageNode.Get("cache_creation.ephemeral_1h_input_tokens").Int()
+	if cc5m > 0 || cc1h > 0 {
+		usage.CacheCreation5mTokens = int(cc5m)
+		usage.CacheCreation1hTokens = int(cc1h)
+	}
+	if usage.CacheCreationInputTokens == 0 && (cc5m > 0 || cc1h > 0) {
+		usage.CacheCreationInputTokens = int(cc5m + cc1h)
+	}
+	if usage.CacheReadInputTokens == 0 {
+		if cached := usageNode.Get("cached_tokens").Int(); cached > 0 {
+			usage.CacheReadInputTokens = int(cached)
 		}
-		return s.handleErrorResponse(ctx, resp, c, account)
 	}
+	return usage
+}
 
-	// 处理正常响应
-	var usage *ClaudeUsage
-	var firstTokenMs *int
-	var clientDisconnect bool
-	if reqStream {
-		streamResult, err := s.handleStreamingResponse(ctx, resp, c, account, startTime, originalModel, reqModel, shouldMimicClaudeCode)
-		if err != nil {
-			if err.Error() == "have error in stream" {
-				return nil, &UpstreamFailoverError{
-					StatusCode: 403,
-				}
-			}
-			return nil, err
-		}
-		usage = streamResult.usage
-		firstTokenMs = streamResult.firstTokenMs
-		clientDisconnect = streamResult.clientDisconnect
-	} else {
-		usage, err = s.handleNonStreamingResponse(ctx, resp, c, account, originalModel, reqModel)
-		if err != nil {
-			return nil, err
+func (s *GatewayService) handleNonStreamingResponseAnthropicAPIKeyPassthrough(
+	ctx context.Context,
+	resp *http.Response,
+	c *gin.Context,
+	account *Account,
+) (*ClaudeUsage, error) {
+	if s.rateLimitService != nil {
+		s.rateLimitService.UpdateSessionWindow(ctx, account, resp.Header)
+	}
+
+	maxBytes := resolveUpstreamResponseReadLimit(s.cfg)
+	body, err := readUpstreamResponseBodyLimited(resp.Body, maxBytes)
+	if err != nil {
+		if errors.Is(err, ErrUpstreamResponseBodyTooLarge) {
+			setOpsUpstreamError(c, http.StatusBadGateway, "upstream response too large", "")
+			c.JSON(http.StatusBadGateway, gin.H{
+				"type": "error",
+				"error": gin.H{
+					"type":    "upstream_error",
+					"message": "Upstream response too large",
+				},
+			})
 		}
+		return nil, err
 	}
 
-	return &ForwardResult{
-		RequestID:        resp.Header.Get("x-request-id"),
-		Usage:            *usage,
-		Model:            originalModel, // 使用原始模型用于计费和日志
-		Stream:           reqStream,
-		Duration:         time.Since(startTime),
-		FirstTokenMs:     firstTokenMs,
-		ClientDisconnect: clientDisconnect,
-	}, nil
+	usage := parseClaudeUsageFromResponseBody(body)
+
+	writeAnthropicPassthroughResponseHeaders(c.Writer.Header(), resp.Header, s.cfg)
+	contentType := strings.TrimSpace(resp.Header.Get("Content-Type"))
+	if contentType == "" {
+		contentType = "application/json"
+	}
+	c.Data(resp.StatusCode, contentType, body)
+	return usage, nil
+}
+
+func writeAnthropicPassthroughResponseHeaders(dst http.Header, src http.Header, cfg *config.Config) {
+	if dst == nil || src == nil {
+		return
+	}
+	if cfg != nil {
+		responseheaders.WriteFilteredHeaders(dst, src, cfg.Security.ResponseHeaders)
+		return
+	}
+	if v := strings.TrimSpace(src.Get("Content-Type")); v != "" {
+		dst.Set("Content-Type", v)
+	}
+	if v := strings.TrimSpace(src.Get("x-request-id")); v != "" {
+		dst.Set("x-request-id", v)
+	}
 }
 
 func (s *GatewayService) buildUpstreamRequest(ctx context.Context, c *gin.Context, account *Account, body []byte, token, tokenType, modelID string, reqStream bool, mimicClaudeCode bool) (*http.Request, error) {
@@ -5082,6 +5682,10 @@ func (s *GatewayService) ForwardCountTokens(ctx context.Context, c *gin.Context,
 		return fmt.Errorf("parse request: empty request")
 	}
 
+	if account != nil && account.IsAnthropicAPIKeyPassthroughEnabled() {
+		return s.forwardCountTokensAnthropicAPIKeyPassthrough(ctx, c, account, parsed.Body)
+	}
+
 	body := parsed.Body
 	reqModel := parsed.Model
 
@@ -5241,6 +5845,158 @@ func (s *GatewayService) ForwardCountTokens(ctx context.Context, c *gin.Context,
 	return nil
 }
 
+func (s *GatewayService) forwardCountTokensAnthropicAPIKeyPassthrough(ctx context.Context, c *gin.Context, account *Account, body []byte) error {
+	token, tokenType, err := s.GetAccessToken(ctx, account)
+	if err != nil {
+		s.countTokensError(c, http.StatusBadGateway, "upstream_error", "Failed to get access token")
+		return err
+	}
+	if tokenType != "apikey" {
+		s.countTokensError(c, http.StatusBadGateway, "upstream_error", "Invalid account token type")
+		return fmt.Errorf("anthropic api key passthrough requires apikey token, got: %s", tokenType)
+	}
+
+	upstreamReq, err := s.buildCountTokensRequestAnthropicAPIKeyPassthrough(ctx, c, account, body, token)
+	if err != nil {
+		s.countTokensError(c, http.StatusInternalServerError, "api_error", "Failed to build request")
+		return err
+	}
+
+	proxyURL := ""
+	if account.ProxyID != nil && account.Proxy != nil {
+		proxyURL = account.Proxy.URL()
+	}
+
+	resp, err := s.httpUpstream.DoWithTLS(upstreamReq, proxyURL, account.ID, account.Concurrency, account.IsTLSFingerprintEnabled())
+	if err != nil {
+		setOpsUpstreamError(c, 0, sanitizeUpstreamErrorMessage(err.Error()), "")
+		appendOpsUpstreamError(c, OpsUpstreamErrorEvent{
+			Platform:           account.Platform,
+			AccountID:          account.ID,
+			AccountName:        account.Name,
+			UpstreamStatusCode: 0,
+			Passthrough:        true,
+			Kind:               "request_error",
+			Message:            sanitizeUpstreamErrorMessage(err.Error()),
+		})
+		s.countTokensError(c, http.StatusBadGateway, "upstream_error", "Request failed")
+		return fmt.Errorf("upstream request failed: %w", err)
+	}
+
+	maxReadBytes := resolveUpstreamResponseReadLimit(s.cfg)
+	respBody, err := readUpstreamResponseBodyLimited(resp.Body, maxReadBytes)
+	_ = resp.Body.Close()
+	if err != nil {
+		if errors.Is(err, ErrUpstreamResponseBodyTooLarge) {
+			setOpsUpstreamError(c, http.StatusBadGateway, "upstream response too large", "")
+			s.countTokensError(c, http.StatusBadGateway, "upstream_error", "Upstream response too large")
+			return err
+		}
+		s.countTokensError(c, http.StatusBadGateway, "upstream_error", "Failed to read response")
+		return err
+	}
+
+	if resp.StatusCode >= 400 {
+		if s.rateLimitService != nil {
+			s.rateLimitService.HandleUpstreamError(ctx, account, resp.StatusCode, resp.Header, respBody)
+		}
+
+		upstreamMsg := strings.TrimSpace(extractUpstreamErrorMessage(respBody))
+		upstreamMsg = sanitizeUpstreamErrorMessage(upstreamMsg)
+		upstreamDetail := ""
+		if s.cfg != nil && s.cfg.Gateway.LogUpstreamErrorBody {
+			maxBytes := s.cfg.Gateway.LogUpstreamErrorBodyMaxBytes
+			if maxBytes <= 0 {
+				maxBytes = 2048
+			}
+			upstreamDetail = truncateString(string(respBody), maxBytes)
+		}
+		setOpsUpstreamError(c, resp.StatusCode, upstreamMsg, upstreamDetail)
+		appendOpsUpstreamError(c, OpsUpstreamErrorEvent{
+			Platform:           account.Platform,
+			AccountID:          account.ID,
+			AccountName:        account.Name,
+			UpstreamStatusCode: resp.StatusCode,
+			UpstreamRequestID:  resp.Header.Get("x-request-id"),
+			Passthrough:        true,
+			Kind:               "http_error",
+			Message:            upstreamMsg,
+			Detail:             upstreamDetail,
+		})
+
+		errMsg := "Upstream request failed"
+		switch resp.StatusCode {
+		case 429:
+			errMsg = "Rate limit exceeded"
+		case 529:
+			errMsg = "Service overloaded"
+		}
+		s.countTokensError(c, resp.StatusCode, "upstream_error", errMsg)
+		if upstreamMsg == "" {
+			return fmt.Errorf("upstream error: %d", resp.StatusCode)
+		}
+		return fmt.Errorf("upstream error: %d message=%s", resp.StatusCode, upstreamMsg)
+	}
+
+	writeAnthropicPassthroughResponseHeaders(c.Writer.Header(), resp.Header, s.cfg)
+	contentType := strings.TrimSpace(resp.Header.Get("Content-Type"))
+	if contentType == "" {
+		contentType = "application/json"
+	}
+	c.Data(resp.StatusCode, contentType, respBody)
+	return nil
+}
+
+func (s *GatewayService) buildCountTokensRequestAnthropicAPIKeyPassthrough(
+	ctx context.Context,
+	c *gin.Context,
+	account *Account,
+	body []byte,
+	token string,
+) (*http.Request, error) {
+	targetURL := claudeAPICountTokensURL
+	baseURL := account.GetBaseURL()
+	if baseURL != "" {
+		validatedURL, err := s.validateUpstreamBaseURL(baseURL)
+		if err != nil {
+			return nil, err
+		}
+		targetURL = validatedURL + "/v1/messages/count_tokens"
+	}
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, targetURL, bytes.NewReader(body))
+	if err != nil {
+		return nil, err
+	}
+
+	if c != nil && c.Request != nil {
+		for key, values := range c.Request.Header {
+			lowerKey := strings.ToLower(strings.TrimSpace(key))
+			if !allowedHeaders[lowerKey] {
+				continue
+			}
+			for _, v := range values {
+				req.Header.Add(key, v)
+			}
+		}
+	}
+
+	req.Header.Del("authorization")
+	req.Header.Del("x-api-key")
+	req.Header.Del("x-goog-api-key")
+	req.Header.Del("cookie")
+	req.Header.Set("x-api-key", token)
+
+	if req.Header.Get("content-type") == "" {
+		req.Header.Set("content-type", "application/json")
+	}
+	if req.Header.Get("anthropic-version") == "" {
+		req.Header.Set("anthropic-version", "2023-06-01")
+	}
+
+	return req, nil
+}
+
 // buildCountTokensRequest 构建 count_tokens 上游请求
 func (s *GatewayService) buildCountTokensRequest(ctx context.Context, c *gin.Context, account *Account, body []byte, token, tokenType, modelID string, mimicClaudeCode bool) (*http.Request, error) {
 	// 确定目标 URL

frontend/src/components/account/CreateAccountModal.vue

@@ -1697,6 +1697,36 @@
         </div>
       </div>
 
+      <!-- Anthropic API Key 自动透传开关 -->
+      <div
+        v-if="form.platform === 'anthropic' && accountCategory === 'apikey'"
+        class="border-t border-gray-200 pt-4 dark:border-dark-600"
+      >
+        <div class="flex items-center justify-between">
+          <div>
+            <label class="input-label mb-0">{{ t('admin.accounts.anthropic.apiKeyPassthrough') }}</label>
+            <p class="mt-1 text-xs text-gray-500 dark:text-gray-400">
+              {{ t('admin.accounts.anthropic.apiKeyPassthroughDesc') }}
+            </p>
+          </div>
+          <button
+            type="button"
+            @click="anthropicPassthroughEnabled = !anthropicPassthroughEnabled"
+            :class="[
+              'relative inline-flex h-6 w-11 flex-shrink-0 cursor-pointer rounded-full border-2 border-transparent transition-colors duration-200 ease-in-out focus:outline-none focus:ring-2 focus:ring-primary-500 focus:ring-offset-2',
+              anthropicPassthroughEnabled ? 'bg-primary-600' : 'bg-gray-200 dark:bg-dark-600'
+            ]"
+          >
+            <span
+              :class="[
+                'pointer-events-none inline-block h-5 w-5 transform rounded-full bg-white shadow ring-0 transition duration-200 ease-in-out',
+                anthropicPassthroughEnabled ? 'translate-x-5' : 'translate-x-0'
+              ]"
+            />
+          </button>
+        </div>
+      </div>
+
       <!-- OpenAI OAuth Codex 官方客户端限制开关 -->
       <div
         v-if="form.platform === 'openai' && accountCategory === 'oauth-based'"
@@ -2290,6 +2320,7 @@ const interceptWarmupRequests = ref(false)
 const autoPauseOnExpired = ref(true)
 const openaiPassthroughEnabled = ref(false)
 const codexCLIOnlyEnabled = ref(false)
+const anthropicPassthroughEnabled = ref(false)
 const mixedScheduling = ref(false) // For antigravity accounts: enable mixed scheduling
 const antigravityAccountType = ref<'oauth' | 'upstream'>('oauth') // For antigravity: oauth or upstream
 const upstreamBaseUrl = ref('') // For upstream type: base URL
@@ -2526,6 +2557,9 @@ watch(
       openaiPassthroughEnabled.value = false
       codexCLIOnlyEnabled.value = false
     }
+    if (newPlatform !== 'anthropic') {
+      anthropicPassthroughEnabled.value = false
+    }
     // Reset OAuth states
     oauth.resetState()
     openaiOAuth.resetState()
@@ -2542,6 +2576,9 @@ watch(
     if (platform === 'openai' && category !== 'oauth-based') {
       codexCLIOnlyEnabled.value = false
     }
+    if (platform !== 'anthropic' || category !== 'apikey') {
+      anthropicPassthroughEnabled.value = false
+    }
   }
 )
 
@@ -2791,6 +2828,7 @@ const resetForm = () => {
   autoPauseOnExpired.value = true
   openaiPassthroughEnabled.value = false
   codexCLIOnlyEnabled.value = false
+  anthropicPassthroughEnabled.value = false
   // Reset quota control state
   windowCostEnabled.value = false
   windowCostLimit.value = null
@@ -2845,6 +2883,21 @@ const buildOpenAIExtra = (base?: Record<string, unknown>): Record<string, unknow
   return Object.keys(extra).length > 0 ? extra : undefined
 }
 
+const buildAnthropicExtra = (base?: Record<string, unknown>): Record<string, unknown> | undefined => {
+  if (form.platform !== 'anthropic' || accountCategory.value !== 'apikey') {
+    return base
+  }
+
+  const extra: Record<string, unknown> = { ...(base || {}) }
+  if (anthropicPassthroughEnabled.value) {
+    extra.anthropic_passthrough = true
+  } else {
+    delete extra.anthropic_passthrough
+  }
+
+  return Object.keys(extra).length > 0 ? extra : undefined
+}
+
 const buildSoraExtra = (
   base?: Record<string, unknown>,
   linkedOpenAIAccountId?: string | number
@@ -3015,7 +3068,7 @@ const handleSubmit = async () => {
   }
 
   form.credentials = credentials
-  const extra = buildOpenAIExtra()
+  const extra = buildAnthropicExtra(buildOpenAIExtra())
 
   await doCreateAccount({
     ...form,

frontend/src/components/account/EditAccountModal.vue

@@ -735,6 +735,36 @@
         </div>
       </div>
 
+      <!-- Anthropic API Key 自动透传开关 -->
+      <div
+        v-if="account?.platform === 'anthropic' && account?.type === 'apikey'"
+        class="border-t border-gray-200 pt-4 dark:border-dark-600"
+      >
+        <div class="flex items-center justify-between">
+          <div>
+            <label class="input-label mb-0">{{ t('admin.accounts.anthropic.apiKeyPassthrough') }}</label>
+            <p class="mt-1 text-xs text-gray-500 dark:text-gray-400">
+              {{ t('admin.accounts.anthropic.apiKeyPassthroughDesc') }}
+            </p>
+          </div>
+          <button
+            type="button"
+            @click="anthropicPassthroughEnabled = !anthropicPassthroughEnabled"
+            :class="[
+              'relative inline-flex h-6 w-11 flex-shrink-0 cursor-pointer rounded-full border-2 border-transparent transition-colors duration-200 ease-in-out focus:outline-none focus:ring-2 focus:ring-primary-500 focus:ring-offset-2',
+              anthropicPassthroughEnabled ? 'bg-primary-600' : 'bg-gray-200 dark:bg-dark-600'
+            ]"
+          >
+            <span
+              :class="[
+                'pointer-events-none inline-block h-5 w-5 transform rounded-full bg-white shadow ring-0 transition duration-200 ease-in-out',
+                anthropicPassthroughEnabled ? 'translate-x-5' : 'translate-x-0'
+              ]"
+            />
+          </button>
+        </div>
+      </div>
+
       <!-- OpenAI OAuth Codex 官方客户端限制开关 -->
       <div
         v-if="account?.platform === 'openai' && account?.type === 'oauth'"
@@ -1223,6 +1253,7 @@ const cacheTTLOverrideTarget = ref<string>('5m')
 // OpenAI 自动透传开关（OAuth/API Key）
 const openaiPassthroughEnabled = ref(false)
 const codexCLIOnlyEnabled = ref(false)
+const anthropicPassthroughEnabled = ref(false)
 const isOpenAIModelRestrictionDisabled = computed(() =>
   props.account?.platform === 'openai' && openaiPassthroughEnabled.value
 )
@@ -1317,12 +1348,16 @@ watch(
       // Load OpenAI passthrough toggle (OpenAI OAuth/API Key)
       openaiPassthroughEnabled.value = false
       codexCLIOnlyEnabled.value = false
+      anthropicPassthroughEnabled.value = false
       if (newAccount.platform === 'openai' && (newAccount.type === 'oauth' || newAccount.type === 'apikey')) {
         openaiPassthroughEnabled.value = extra?.openai_passthrough === true || extra?.openai_oauth_passthrough === true
         if (newAccount.type === 'oauth') {
           codexCLIOnlyEnabled.value = extra?.codex_cli_only === true
         }
       }
+      if (newAccount.platform === 'anthropic' && newAccount.type === 'apikey') {
+        anthropicPassthroughEnabled.value = extra?.anthropic_passthrough === true
+      }
 
       // Load antigravity model mapping (Antigravity 只支持映射模式)
       if (newAccount.platform === 'antigravity') {
@@ -1882,6 +1917,18 @@ const handleSubmit = async () => {
       updatePayload.extra = newExtra
     }
 
+    // For Anthropic API Key accounts, handle passthrough mode in extra
+    if (props.account.platform === 'anthropic' && props.account.type === 'apikey') {
+      const currentExtra = (props.account.extra as Record<string, unknown>) || {}
+      const newExtra: Record<string, unknown> = { ...currentExtra }
+      if (anthropicPassthroughEnabled.value) {
+        newExtra.anthropic_passthrough = true
+      } else {
+        delete newExtra.anthropic_passthrough
+      }
+      updatePayload.extra = newExtra
+    }
+
     // For OpenAI OAuth/API Key accounts, handle passthrough mode in extra
     if (props.account.platform === 'openai' && (props.account.type === 'oauth' || props.account.type === 'apikey')) {
       const currentExtra = (props.account.extra as Record<string, unknown>) || {}

frontend/src/i18n/locales/en.ts

@@ -1548,6 +1548,11 @@ export default {
         enableSora: 'Enable Sora simultaneously',
         enableSoraHint: 'Sora uses the same OpenAI account. Enable to create Sora account simultaneously.'
       },
+      anthropic: {
+        apiKeyPassthrough: 'Auto passthrough (auth only)',
+        apiKeyPassthroughDesc:
+          'Only applies to Anthropic API Key accounts. When enabled, messages/count_tokens are forwarded in passthrough mode with auth replacement only, while billing/concurrency/audit and safety filtering are preserved. Disable to roll back immediately.'
+      },
       modelRestriction: 'Model Restriction (Optional)',
       modelWhitelist: 'Model Whitelist',
       modelMapping: 'Model Mapping',

frontend/src/i18n/locales/zh.ts

@@ -1696,6 +1696,11 @@ export default {
         enableSora: '同时启用 Sora',
         enableSoraHint: 'Sora 使用相同的 OpenAI 账号，开启后将同时创建 Sora 平台账号'
       },
+      anthropic: {
+        apiKeyPassthrough: '自动透传（仅替换认证）',
+        apiKeyPassthroughDesc:
+          '仅对 Anthropic API Key 生效。开启后，messages/count_tokens 请求将透传上游并仅替换认证，保留计费/并发/审计及必要安全过滤；关闭即可回滚到现有兼容链路。'
+      },
       modelRestriction: '模型限制（可选）',
       modelWhitelist: '模型白名单',
       modelMapping: '模型映射',