diff --git a/backend/internal/service/gateway_forward_as_chat_completions.go b/backend/internal/service/gateway_forward_as_chat_completions.go
index 93da85a7b2..d03cf53531 100644
--- a/backend/internal/service/gateway_forward_as_chat_completions.go
+++ b/backend/internal/service/gateway_forward_as_chat_completions.go
@@ -179,7 +179,7 @@ func (s *GatewayService) ForwardAsChatCompletions(
 	}
 
 	// 13. Extract reasoning effort from CC request body
-	reasoningEffort := extractCCReasoningEffortFromBody(body)
+	reasoningEffort := extractCCReasoningEffortFromBody(body, mappedModel, originalModel)
 	// 国产模型默认 effort 补充：本路径是客户端 CC 请求 → Anthropic 上游，
 	// 如果上游是 passback-required 国产模型 (Kimi-anthropic / GLM-anthropic / MiniMax)
 	// 且客户端在 body 里传了 thinking.type=enabled，补中默认 effort。
@@ -201,7 +201,7 @@ func (s *GatewayService) ForwardAsChatCompletions(
 // extractCCReasoningEffortFromBody reads reasoning effort from a Chat Completions
 // request body. It checks both nested (reasoning.effort) and flat (reasoning_effort)
 // formats used by OpenAI-compatible clients.
-func extractCCReasoningEffortFromBody(body []byte) *string {
+func extractCCReasoningEffortFromBody(body []byte, modelCandidates ...string) *string {
 	raw := strings.TrimSpace(gjson.GetBytes(body, "reasoning.effort").String())
 	if raw == "" {
 		raw = strings.TrimSpace(gjson.GetBytes(body, "reasoning_effort").String())
@@ -209,7 +209,11 @@ func extractCCReasoningEffortFromBody(body []byte) *string {
 	if raw == "" {
 		return nil
 	}
-	normalized := normalizeOpenAIReasoningEffort(raw)
+	model := firstNonEmpty(modelCandidates...)
+	if model == "" {
+		model = strings.TrimSpace(gjson.GetBytes(body, "model").String())
+	}
+	normalized := normalizeOpenAIReasoningEffortForModel(raw, model)
 	if normalized == "" {
 		return nil
 	}
diff --git a/backend/internal/service/gateway_forward_as_chat_completions_test.go b/backend/internal/service/gateway_forward_as_chat_completions_test.go
index d1047bdc9b..a5a5cbf913 100644
--- a/backend/internal/service/gateway_forward_as_chat_completions_test.go
+++ b/backend/internal/service/gateway_forward_as_chat_completions_test.go
@@ -30,7 +30,23 @@ func TestExtractCCReasoningEffortFromBody(t *testing.T) {
 	})
 
 	t.Run("DeepSeek max", func(t *testing.T) {
-		got := extractCCReasoningEffortFromBody([]byte(`{"reasoning_effort":"Max"}`))
+		got := extractCCReasoningEffortFromBody([]byte(`{"model":"deepseek-v4-flash","reasoning_effort":"Max"}`))
+		require.NotNil(t, got)
+		require.Equal(t, "max", *got)
+	})
+
+	t.Run("mapped Kimi alias max", func(t *testing.T) {
+		got := extractCCReasoningEffortFromBody(
+			[]byte(`{"model":"public-alias","reasoning_effort":"max"}`),
+			"kimi-k3",
+			"public-alias",
+		)
+		require.NotNil(t, got)
+		require.Equal(t, "max", *got)
+	})
+
+	t.Run("legacy model max", func(t *testing.T) {
+		got := extractCCReasoningEffortFromBody([]byte(`{"model":"gpt-5.5","reasoning_effort":"max"}`))
 		require.NotNil(t, got)
 		require.Equal(t, "xhigh", *got)
 	})
diff --git a/backend/internal/service/gateway_forward_as_responses.go b/backend/internal/service/gateway_forward_as_responses.go
index 2aa5255c29..b0ca29c1c2 100644
--- a/backend/internal/service/gateway_forward_as_responses.go
+++ b/backend/internal/service/gateway_forward_as_responses.go
@@ -63,7 +63,6 @@ func (s *GatewayService) ForwardAsResponses(
 
 	// 4. Model mapping
 	mappedModel := originalModel
-	reasoningEffort := ExtractResponsesReasoningEffortFromBody(body)
 	if account.Type == AccountTypeAPIKey || account.Type == AccountTypeServiceAccount {
 		mappedModel = account.GetMappedModel(originalModel)
 	}
@@ -78,6 +77,7 @@ func (s *GatewayService) ForwardAsResponses(
 			mappedModel = normalized
 		}
 	}
+	reasoningEffort := ExtractResponsesReasoningEffortFromBody(body, mappedModel, originalModel)
 	// 国产模型默认 effort 补充：需要 mappedModel 判定，推迟到 mapping 完成之后。
 	reasoningEffort = ApplyThinkingEnabledFallback(reasoningEffort, body, mappedModel)
 	anthropicReq.Model = mappedModel
@@ -255,12 +255,16 @@ func liftResponsesAdditionalTools(requestBody map[string]any) (bool, error) {
 
 // ExtractResponsesReasoningEffortFromBody reads Responses API reasoning.effort
 // and normalizes it for usage logging.
-func ExtractResponsesReasoningEffortFromBody(body []byte) *string {
+func ExtractResponsesReasoningEffortFromBody(body []byte, modelCandidates ...string) *string {
 	raw := strings.TrimSpace(gjson.GetBytes(body, "reasoning.effort").String())
 	if raw == "" {
 		return nil
 	}
-	normalized := normalizeOpenAIReasoningEffort(raw)
+	model := firstNonEmpty(modelCandidates...)
+	if model == "" {
+		model = strings.TrimSpace(gjson.GetBytes(body, "model").String())
+	}
+	normalized := normalizeOpenAIReasoningEffortForModel(raw, model)
 	if normalized == "" {
 		return nil
 	}
diff --git a/backend/internal/service/gateway_forward_as_responses_test.go b/backend/internal/service/gateway_forward_as_responses_test.go
index 6784e7c7a5..262d588629 100644
--- a/backend/internal/service/gateway_forward_as_responses_test.go
+++ b/backend/internal/service/gateway_forward_as_responses_test.go
@@ -143,7 +143,19 @@ func TestExtractResponsesReasoningEffortFromBody(t *testing.T) {
 
 	maxGot := ExtractResponsesReasoningEffortFromBody([]byte(`{"model":"deepseek-v4-pro","reasoning":{"effort":"max"}}`))
 	require.NotNil(t, maxGot)
-	require.Equal(t, "xhigh", *maxGot)
+	require.Equal(t, "max", *maxGot)
+
+	mappedMax := ExtractResponsesReasoningEffortFromBody(
+		[]byte(`{"model":"public-alias","reasoning":{"effort":"max"}}`),
+		"provider/glm-5.2",
+		"public-alias",
+	)
+	require.NotNil(t, mappedMax)
+	require.Equal(t, "max", *mappedMax)
+
+	legacyMax := ExtractResponsesReasoningEffortFromBody([]byte(`{"model":"gpt-5.5","reasoning":{"effort":"max"}}`))
+	require.NotNil(t, legacyMax)
+	require.Equal(t, "xhigh", *legacyMax)
 
 	require.Nil(t, ExtractResponsesReasoningEffortFromBody([]byte(`{"model":"claude-sonnet-4.5"}`)))
 }
diff --git a/backend/internal/service/gemini_chat_completions_compat_service.go b/backend/internal/service/gemini_chat_completions_compat_service.go
index 3ee7a4af9e..a56a1b3ebc 100644
--- a/backend/internal/service/gemini_chat_completions_compat_service.go
+++ b/backend/internal/service/gemini_chat_completions_compat_service.go
@@ -204,7 +204,7 @@ func (s *GeminiMessagesCompatService) forwardClaudeBodyAsChatCompletions(
 		c.Header("x-request-id", requestID)
 	}
 
-	reasoningEffort := extractCCReasoningEffortFromBody(originalChatBody)
+	reasoningEffort := extractCCReasoningEffortFromBody(originalChatBody, mappedModel)
 	// 国产模型默认 effort 补充（本路径上游是 Gemini，不会命中 passback-required）。
 	// 保持与 OpenAI 网关路径调用模式一致，便于未来上游变异时语义一致。
 	reasoningEffort = ApplyThinkingEnabledFallback(reasoningEffort, originalChatBody, mappedModel)
diff --git a/backend/internal/service/openai_gateway_chat_completions_anthropic_native.go b/backend/internal/service/openai_gateway_chat_completions_anthropic_native.go
index b41bc644c6..d4000f9289 100644
--- a/backend/internal/service/openai_gateway_chat_completions_anthropic_native.go
+++ b/backend/internal/service/openai_gateway_chat_completions_anthropic_native.go
@@ -132,7 +132,7 @@ func (s *OpenAIGatewayService) forwardChatCompletionsViaNativeAnthropic(
 		return nil, fmt.Errorf("upstream error: %d %s", resp.StatusCode, upstreamMsg)
 	}
 
-	reasoningEffort := extractCCReasoningEffortFromBody(body)
+	reasoningEffort := extractCCReasoningEffortFromBody(body, upstreamModel, billingModel, originalModel)
 	reasoningEffort = ApplyThinkingEnabledFallback(reasoningEffort, body, billingModel)
 
 	if clientStream {
diff --git a/backend/internal/service/openai_gateway_request_body.go b/backend/internal/service/openai_gateway_request_body.go
index 59199cea05..ad9b6abaea 100644
--- a/backend/internal/service/openai_gateway_request_body.go
+++ b/backend/internal/service/openai_gateway_request_body.go
@@ -1507,8 +1507,31 @@ func normalizeOpenAIReasoningEffort(raw string) string {
 }
 
 func normalizeOpenAIReasoningEffortForModel(raw, model string) string {
-	if strings.EqualFold(strings.TrimSpace(raw), "max") && isOpenAIGPT56Model(model) {
+	if strings.EqualFold(strings.TrimSpace(raw), "max") && supportsOpenAIReasoningEffortMax(model) {
 		return "max"
 	}
 	return normalizeOpenAIReasoningEffort(raw)
 }
+
+// supportsOpenAIReasoningEffortMax reports model families whose upstream scale
+// has a distinct max level. Other models keep the legacy max -> xhigh behavior.
+func supportsOpenAIReasoningEffortMax(model string) bool {
+	if isOpenAIGPT56Model(model) {
+		return true
+	}
+
+	normalized := strings.ToLower(lastOpenAIModelSegment(model))
+	normalized = strings.ReplaceAll(normalized, "_", "-")
+	switch {
+	case strings.HasPrefix(normalized, "deepseek-v4"):
+		return true
+	case strings.HasPrefix(normalized, "glm-"):
+		return true
+	case strings.HasPrefix(normalized, "kimi-"), strings.HasPrefix(normalized, "moonshot-"):
+		return true
+	case normalized == "k3" || strings.HasPrefix(normalized, "k3-"):
+		return true
+	default:
+		return false
+	}
+}
diff --git a/backend/internal/service/openai_gateway_responses_anthropic_native.go b/backend/internal/service/openai_gateway_responses_anthropic_native.go
index 04cd7fc274..88cf83310a 100644
--- a/backend/internal/service/openai_gateway_responses_anthropic_native.go
+++ b/backend/internal/service/openai_gateway_responses_anthropic_native.go
@@ -75,7 +75,7 @@ func (s *OpenAIGatewayService) forwardResponsesViaNativeAnthropic(
 	upstreamModel := normalizeOpenAIModelForUpstream(account, billingModel)
 	anthropicReq.Model = upstreamModel
 
-	reasoningEffort := ExtractResponsesReasoningEffortFromBody(body)
+	reasoningEffort := ExtractResponsesReasoningEffortFromBody(body, upstreamModel, billingModel, originalModel)
 	reasoningEffort = ApplyThinkingEnabledFallback(reasoningEffort, body, billingModel)
 
 	// 5. Force upstream streaming（客户端原始终决定响应格式；
diff --git a/backend/internal/service/openai_gateway_service_hotpath_test.go b/backend/internal/service/openai_gateway_service_hotpath_test.go
index a51f58bf56..b6c3abca35 100644
--- a/backend/internal/service/openai_gateway_service_hotpath_test.go
+++ b/backend/internal/service/openai_gateway_service_hotpath_test.go
@@ -920,10 +920,17 @@ func TestExtractOpenAIReasoningEffortFromBody(t *testing.T) {
 			wantValue: "xhigh",
 		},
 		{
-			name:      "DeepSeek max 归一化为 xhigh",
+			name:      "DeepSeek V4 保留 max",
 			body:      []byte(`{"reasoning_effort":"max"}`),
 			model:     "deepseek-v4-pro",
 			wantNil:   false,
+			wantValue: "max",
+		},
+		{
+			name:      "旧模型仍将 max 归一化为 xhigh",
+			body:      []byte(`{"reasoning_effort":"max"}`),
+			model:     "gpt-5.5",
+			wantNil:   false,
 			wantValue: "xhigh",
 		},
 		{
diff --git a/backend/internal/service/openai_gpt56_max_test.go b/backend/internal/service/openai_gpt56_max_test.go
index cbca2ff3ee..9dc865295c 100644
--- a/backend/internal/service/openai_gpt56_max_test.go
+++ b/backend/internal/service/openai_gpt56_max_test.go
@@ -24,7 +24,8 @@ func TestNormalizeOpenAIReasoningEffortForGPT56(t *testing.T) {
 		{name: "Sol 保留 max", raw: "max", model: "gpt-5.6-sol", want: "max"},
 		{name: "Terra 保留 max", raw: "max", model: "openai/gpt-5.6-terra", want: "max"},
 		{name: "Luna 后缀保留 max", raw: "max", model: "gpt-5.6-luna-2026-07-09", want: "max"},
-		{name: "其他模型沿用 xhigh", raw: "max", model: "deepseek-v4-pro", want: "xhigh"},
+		{name: "DeepSeek V4 保留 max", raw: "max", model: "deepseek-v4-pro", want: "max"},
+		{name: "旧 GPT 模型沿用 xhigh", raw: "max", model: "gpt-5.5", want: "xhigh"},
 	}
 
 	for _, tt := range tests {