Merge pull request #788 from MartialBE/main

feat: Add Claude 3.7 Sonnet thinking mode support

common/model-ratio.go

@@ -83,93 +83,94 @@ var defaultModelRatio = map[string]float64{
 	"text-curie-001":         1,
 	//"text-davinci-002":               10,
 	//"text-davinci-003":               10,
-	"text-davinci-edit-001":          10,
-	"code-davinci-edit-001":          10,
-	"whisper-1":                      15,  // $0.006 / minute -> $0.006 / 150 words -> $0.006 / 200 tokens -> $0.03 / 1k tokens
-	"tts-1":                          7.5, // 1k characters -> $0.015
-	"tts-1-1106":                     7.5, // 1k characters -> $0.015
-	"tts-1-hd":                       15,  // 1k characters -> $0.03
-	"tts-1-hd-1106":                  15,  // 1k characters -> $0.03
-	"davinci":                        10,
-	"curie":                          10,
-	"babbage":                        10,
-	"ada":                            10,
-	"text-embedding-3-small":         0.01,
-	"text-embedding-3-large":         0.065,
-	"text-embedding-ada-002":         0.05,
-	"text-search-ada-doc-001":        10,
-	"text-moderation-stable":         0.1,
-	"text-moderation-latest":         0.1,
-	"claude-instant-1":               0.4,   // $0.8 / 1M tokens
-	"claude-2.0":                     4,     // $8 / 1M tokens
-	"claude-2.1":                     4,     // $8 / 1M tokens
-	"claude-3-haiku-20240307":        0.125, // $0.25 / 1M tokens
-	"claude-3-5-haiku-20241022":      0.5,   // $1 / 1M tokens
-	"claude-3-sonnet-20240229":       1.5,   // $3 / 1M tokens
-	"claude-3-5-sonnet-20240620":     1.5,
-	"claude-3-5-sonnet-20241022":     1.5,
-	"claude-3-7-sonnet-20250219":     1.5,
-	"claude-3-opus-20240229":         7.5, // $15 / 1M tokens
-	"ERNIE-4.0-8K":                   0.120 * RMB,
-	"ERNIE-3.5-8K":                   0.012 * RMB,
-	"ERNIE-3.5-8K-0205":              0.024 * RMB,
-	"ERNIE-3.5-8K-1222":              0.012 * RMB,
-	"ERNIE-Bot-8K":                   0.024 * RMB,
-	"ERNIE-3.5-4K-0205":              0.012 * RMB,
-	"ERNIE-Speed-8K":                 0.004 * RMB,
-	"ERNIE-Speed-128K":               0.004 * RMB,
-	"ERNIE-Lite-8K-0922":             0.008 * RMB,
-	"ERNIE-Lite-8K-0308":             0.003 * RMB,
-	"ERNIE-Tiny-8K":                  0.001 * RMB,
-	"BLOOMZ-7B":                      0.004 * RMB,
-	"Embedding-V1":                   0.002 * RMB,
-	"bge-large-zh":                   0.002 * RMB,
-	"bge-large-en":                   0.002 * RMB,
-	"tao-8k":                         0.002 * RMB,
-	"PaLM-2":                         1,
-	"gemini-pro":                     1, // $0.00025 / 1k characters -> $0.001 / 1k tokens
-	"gemini-pro-vision":              1, // $0.00025 / 1k characters -> $0.001 / 1k tokens
-	"gemini-1.0-pro-vision-001":      1,
-	"gemini-1.0-pro-001":             1,
-	"gemini-1.5-pro-latest":          1.75, // $3.5 / 1M tokens
-	"gemini-1.5-pro-exp-0827":        1.75, // $3.5 / 1M tokens
-	"gemini-1.5-flash-latest":        1,
-	"gemini-1.5-flash-exp-0827":      1,
-	"gemini-1.0-pro-latest":          1,
-	"gemini-1.0-pro-vision-latest":   1,
-	"gemini-ultra":                   1,
-	"chatglm_turbo":                  0.3572,     // ￥0.005 / 1k tokens
-	"chatglm_pro":                    0.7143,     // ￥0.01 / 1k tokens
-	"chatglm_std":                    0.3572,     // ￥0.005 / 1k tokens
-	"chatglm_lite":                   0.1429,     // ￥0.002 / 1k tokens
-	"glm-4":                          7.143,      // ￥0.1 / 1k tokens
-	"glm-4v":                         0.05 * RMB, // ￥0.05 / 1k tokens
-	"glm-4-alltools":                 0.1 * RMB,  // ￥0.1 / 1k tokens
-	"glm-3-turbo":                    0.3572,
-	"glm-4-plus":                     0.05 * RMB,
-	"glm-4-0520":                     0.1 * RMB,
-	"glm-4-air":                      0.001 * RMB,
-	"glm-4-airx":                     0.01 * RMB,
-	"glm-4-long":                     0.001 * RMB,
-	"glm-4-flash":                    0,
-	"glm-4v-plus":                    0.01 * RMB,
-	"qwen-turbo":                     0.8572, // ￥0.012 / 1k tokens
-	"qwen-plus":                      10,     // ￥0.14 / 1k tokens
-	"text-embedding-v1":              0.05,   // ￥0.0007 / 1k tokens
-	"SparkDesk-v1.1":                 1.2858, // ￥0.018 / 1k tokens
-	"SparkDesk-v2.1":                 1.2858, // ￥0.018 / 1k tokens
-	"SparkDesk-v3.1":                 1.2858, // ￥0.018 / 1k tokens
-	"SparkDesk-v3.5":                 1.2858, // ￥0.018 / 1k tokens
-	"SparkDesk-v4.0":                 1.2858,
-	"360GPT_S2_V9":                   0.8572, // ¥0.012 / 1k tokens
-	"360gpt-turbo":                   0.0858, // ¥0.0012 / 1k tokens
-	"360gpt-turbo-responsibility-8k": 0.8572, // ¥0.012 / 1k tokens
-	"360gpt-pro":                     0.8572, // ¥0.012 / 1k tokens
-	"360gpt2-pro":                    0.8572, // ¥0.012 / 1k tokens
-	"embedding-bert-512-v1":          0.0715, // ¥0.001 / 1k tokens
-	"embedding_s1_v1":                0.0715, // ¥0.001 / 1k tokens
-	"semantic_similarity_s1_v1":      0.0715, // ¥0.001 / 1k tokens
-	"hunyuan":                        7.143,  // ¥0.1 / 1k tokens  // https://cloud.tencent.com/document/product/1729/97731#e0e6be58-60c8-469f-bdeb-6c264ce3b4d0
+	"text-davinci-edit-001":               10,
+	"code-davinci-edit-001":               10,
+	"whisper-1":                           15,  // $0.006 / minute -> $0.006 / 150 words -> $0.006 / 200 tokens -> $0.03 / 1k tokens
+	"tts-1":                               7.5, // 1k characters -> $0.015
+	"tts-1-1106":                          7.5, // 1k characters -> $0.015
+	"tts-1-hd":                            15,  // 1k characters -> $0.03
+	"tts-1-hd-1106":                       15,  // 1k characters -> $0.03
+	"davinci":                             10,
+	"curie":                               10,
+	"babbage":                             10,
+	"ada":                                 10,
+	"text-embedding-3-small":              0.01,
+	"text-embedding-3-large":              0.065,
+	"text-embedding-ada-002":              0.05,
+	"text-search-ada-doc-001":             10,
+	"text-moderation-stable":              0.1,
+	"text-moderation-latest":              0.1,
+	"claude-instant-1":                    0.4,   // $0.8 / 1M tokens
+	"claude-2.0":                          4,     // $8 / 1M tokens
+	"claude-2.1":                          4,     // $8 / 1M tokens
+	"claude-3-haiku-20240307":             0.125, // $0.25 / 1M tokens
+	"claude-3-5-haiku-20241022":           0.5,   // $1 / 1M tokens
+	"claude-3-sonnet-20240229":            1.5,   // $3 / 1M tokens
+	"claude-3-5-sonnet-20240620":          1.5,
+	"claude-3-5-sonnet-20241022":          1.5,
+	"claude-3-7-sonnet-20250219":          1.5,
+	"claude-3-7-sonnet-20250219-thinking": 1.5,
+	"claude-3-opus-20240229":              7.5, // $15 / 1M tokens
+	"ERNIE-4.0-8K":                        0.120 * RMB,
+	"ERNIE-3.5-8K":                        0.012 * RMB,
+	"ERNIE-3.5-8K-0205":                   0.024 * RMB,
+	"ERNIE-3.5-8K-1222":                   0.012 * RMB,
+	"ERNIE-Bot-8K":                        0.024 * RMB,
+	"ERNIE-3.5-4K-0205":                   0.012 * RMB,
+	"ERNIE-Speed-8K":                      0.004 * RMB,
+	"ERNIE-Speed-128K":                    0.004 * RMB,
+	"ERNIE-Lite-8K-0922":                  0.008 * RMB,
+	"ERNIE-Lite-8K-0308":                  0.003 * RMB,
+	"ERNIE-Tiny-8K":                       0.001 * RMB,
+	"BLOOMZ-7B":                           0.004 * RMB,
+	"Embedding-V1":                        0.002 * RMB,
+	"bge-large-zh":                        0.002 * RMB,
+	"bge-large-en":                        0.002 * RMB,
+	"tao-8k":                              0.002 * RMB,
+	"PaLM-2":                              1,
+	"gemini-pro":                          1, // $0.00025 / 1k characters -> $0.001 / 1k tokens
+	"gemini-pro-vision":                   1, // $0.00025 / 1k characters -> $0.001 / 1k tokens
+	"gemini-1.0-pro-vision-001":           1,
+	"gemini-1.0-pro-001":                  1,
+	"gemini-1.5-pro-latest":               1.75, // $3.5 / 1M tokens
+	"gemini-1.5-pro-exp-0827":             1.75, // $3.5 / 1M tokens
+	"gemini-1.5-flash-latest":             1,
+	"gemini-1.5-flash-exp-0827":           1,
+	"gemini-1.0-pro-latest":               1,
+	"gemini-1.0-pro-vision-latest":        1,
+	"gemini-ultra":                        1,
+	"chatglm_turbo":                       0.3572,     // ￥0.005 / 1k tokens
+	"chatglm_pro":                         0.7143,     // ￥0.01 / 1k tokens
+	"chatglm_std":                         0.3572,     // ￥0.005 / 1k tokens
+	"chatglm_lite":                        0.1429,     // ￥0.002 / 1k tokens
+	"glm-4":                               7.143,      // ￥0.1 / 1k tokens
+	"glm-4v":                              0.05 * RMB, // ￥0.05 / 1k tokens
+	"glm-4-alltools":                      0.1 * RMB,  // ￥0.1 / 1k tokens
+	"glm-3-turbo":                         0.3572,
+	"glm-4-plus":                          0.05 * RMB,
+	"glm-4-0520":                          0.1 * RMB,
+	"glm-4-air":                           0.001 * RMB,
+	"glm-4-airx":                          0.01 * RMB,
+	"glm-4-long":                          0.001 * RMB,
+	"glm-4-flash":                         0,
+	"glm-4v-plus":                         0.01 * RMB,
+	"qwen-turbo":                          0.8572, // ￥0.012 / 1k tokens
+	"qwen-plus":                           10,     // ￥0.14 / 1k tokens
+	"text-embedding-v1":                   0.05,   // ￥0.0007 / 1k tokens
+	"SparkDesk-v1.1":                      1.2858, // ￥0.018 / 1k tokens
+	"SparkDesk-v2.1":                      1.2858, // ￥0.018 / 1k tokens
+	"SparkDesk-v3.1":                      1.2858, // ￥0.018 / 1k tokens
+	"SparkDesk-v3.5":                      1.2858, // ￥0.018 / 1k tokens
+	"SparkDesk-v4.0":                      1.2858,
+	"360GPT_S2_V9":                        0.8572, // ¥0.012 / 1k tokens
+	"360gpt-turbo":                        0.0858, // ¥0.0012 / 1k tokens
+	"360gpt-turbo-responsibility-8k":      0.8572, // ¥0.012 / 1k tokens
+	"360gpt-pro":                          0.8572, // ¥0.012 / 1k tokens
+	"360gpt2-pro":                         0.8572, // ¥0.012 / 1k tokens
+	"embedding-bert-512-v1":               0.0715, // ¥0.001 / 1k tokens
+	"embedding_s1_v1":                     0.0715, // ¥0.001 / 1k tokens
+	"semantic_similarity_s1_v1":           0.0715, // ¥0.001 / 1k tokens
+	"hunyuan":                             7.143,  // ¥0.1 / 1k tokens  // https://cloud.tencent.com/document/product/1729/97731#e0e6be58-60c8-469f-bdeb-6c264ce3b4d0
 	// https://platform.lingyiwanwu.com/docs#-计费单元
 	// 已经按照 7.2 来换算美元价格
 	"yi-34b-chat-0205":       0.18,

relay/channel/claude/constants.go

@@ -12,6 +12,7 @@ var ModelList = []string{
 	"claude-3-5-sonnet-20240620",
 	"claude-3-5-sonnet-20241022",
 	"claude-3-7-sonnet-20250219",
+	"claude-3-7-sonnet-20250219-thinking",
 }
 
 var ChannelName = "claude"

relay/channel/claude/dto.go

@@ -11,6 +11,9 @@ type ClaudeMediaMessage struct {
 	Usage       *ClaudeUsage         `json:"usage,omitempty"`
 	StopReason  *string              `json:"stop_reason,omitempty"`
 	PartialJson string               `json:"partial_json,omitempty"`
+	Thinking    string               `json:"thinking,omitempty"`
+	Signature   string               `json:"signature,omitempty"`
+	Delta       string               `json:"delta,omitempty"`
 	// tool_calls
 	Id        string `json:"id,omitempty"`
 	Name      string `json:"name,omitempty"`

relay/channel/claude/relay-claude.go

@@ -92,6 +92,25 @@ func RequestOpenAI2ClaudeMessage(textRequest dto.GeneralOpenAIRequest) (*ClaudeR
 		Stream:        textRequest.Stream,
 		Tools:         claudeTools,
 	}
+
+	if strings.HasSuffix(textRequest.Model, "-thinking") {
+		if claudeRequest.MaxTokens == 0 {
+			claudeRequest.MaxTokens = 8192
+		}
+
+		// 因为BudgetTokens 必须大于1024
+		if claudeRequest.MaxTokens < 1280 {
+			claudeRequest.MaxTokens = 1280
+		}
+
+		// BudgetTokens 为 max_tokens 的 80%
+		claudeRequest.Thinking = &Thinking{
+			Type:         "enabled",
+			BudgetTokens: int(float64(claudeRequest.MaxTokens) * 0.8),
+		}
+		claudeRequest.Model = strings.TrimSuffix(textRequest.Model, "-thinking")
+	}
+
 	if claudeRequest.MaxTokens == 0 {
 		claudeRequest.MaxTokens = 4096
 	}
@@ -308,12 +327,20 @@ func StreamResponseClaude2OpenAI(reqMode int, claudeResponse *ClaudeResponse) (*
 			if claudeResponse.Delta != nil {
 				choice.Index = claudeResponse.Index
 				choice.Delta.SetContentString(claudeResponse.Delta.Text)
-				if claudeResponse.Delta.Type == "input_json_delta" {
+				switch claudeResponse.Delta.Type {
+				case "input_json_delta":
 					tools = append(tools, dto.ToolCall{
 						Function: dto.FunctionCall{
 							Arguments: claudeResponse.Delta.PartialJson,
 						},
 					})
+				case "signature_delta":
+					// 加密的不处理
+					signatureContent := "\n"
+					choice.Delta.ReasoningContent = &signatureContent
+				case "thinking_delta":
+					thinkingContent := claudeResponse.Delta.Thinking
+					choice.Delta.ReasoningContent = &thinkingContent
 				}
 			}
 		} else if claudeResponse.Type == "message_delta" {
@@ -352,6 +379,8 @@ func ResponseClaude2OpenAI(reqMode int, claudeResponse *ClaudeResponse) *dto.Ope
 		responseText = claudeResponse.Content[0].Text
 	}
 	tools := make([]dto.ToolCall, 0)
+	thinkingContent := ""
+
 	if reqMode == RequestModeCompletion {
 		content, _ := json.Marshal(strings.TrimPrefix(claudeResponse.Completion, " "))
 		choice := dto.OpenAITextResponseChoice{
@@ -367,7 +396,8 @@ func ResponseClaude2OpenAI(reqMode int, claudeResponse *ClaudeResponse) *dto.Ope
 	} else {
 		fullTextResponse.Id = claudeResponse.Id
 		for _, message := range claudeResponse.Content {
-			if message.Type == "tool_use" {
+			switch message.Type {
+			case "tool_use":
 				args, _ := json.Marshal(message.Input)
 				tools = append(tools, dto.ToolCall{
 					ID:   message.Id,
@@ -377,6 +407,11 @@ func ResponseClaude2OpenAI(reqMode int, claudeResponse *ClaudeResponse) *dto.Ope
 						Arguments: string(args),
 					},
 				})
+			case "thinking":
+				// 加密的不管， 只输出明文的推理过程
+				thinkingContent = message.Thinking
+			case "text":
+				responseText = message.Text
 			}
 		}
 	}
@@ -391,6 +426,7 @@ func ResponseClaude2OpenAI(reqMode int, claudeResponse *ClaudeResponse) *dto.Ope
 	if len(tools) > 0 {
 		choice.Message.SetToolCalls(tools)
 	}
+	choice.Message.ReasoningContent = thinkingContent
 	fullTextResponse.Model = claudeResponse.Model
 	choices = append(choices, choice)
 	fullTextResponse.Choices = choices