Merge pull request #1970 from deqiying/fix-1754-claude-openai-cache-usage

fix(anthropic): 修正缓存 token 的 Anthropic 用量语义

backend/internal/pkg/apicompat/anthropic_responses_test.go

@@ -181,6 +181,55 @@ func TestResponsesToAnthropic_TextOnly(t *testing.T) {
 	assert.Equal(t, 5, anth.Usage.OutputTokens)
 }
 
+func TestResponsesToAnthropic_CachedTokensUseAnthropicInputSemantics(t *testing.T) {
+	resp := &ResponsesResponse{
+		ID:     "resp_cached",
+		Model:  "gpt-5.2",
+		Status: "completed",
+		Output: []ResponsesOutput{
+			{
+				Type: "message",
+				Content: []ResponsesContentPart{
+					{Type: "output_text", Text: "Cached response"},
+				},
+			},
+		},
+		Usage: &ResponsesUsage{
+			InputTokens:  54006,
+			OutputTokens: 123,
+			TotalTokens:  54129,
+			InputTokensDetails: &ResponsesInputTokensDetails{
+				CachedTokens: 50688,
+			},
+		},
+	}
+
+	anth := ResponsesToAnthropic(resp, "claude-sonnet-4-5-20250929")
+	assert.Equal(t, 3318, anth.Usage.InputTokens)
+	assert.Equal(t, 50688, anth.Usage.CacheReadInputTokens)
+	assert.Equal(t, 123, anth.Usage.OutputTokens)
+}
+
+func TestResponsesToAnthropic_CachedTokensClampInputTokens(t *testing.T) {
+	resp := &ResponsesResponse{
+		ID:     "resp_cached_clamp",
+		Model:  "gpt-5.2",
+		Status: "completed",
+		Usage: &ResponsesUsage{
+			InputTokens:  100,
+			OutputTokens: 5,
+			InputTokensDetails: &ResponsesInputTokensDetails{
+				CachedTokens: 150,
+			},
+		},
+	}
+
+	anth := ResponsesToAnthropic(resp, "claude-sonnet-4-5-20250929")
+	assert.Equal(t, 0, anth.Usage.InputTokens)
+	assert.Equal(t, 150, anth.Usage.CacheReadInputTokens)
+	assert.Equal(t, 5, anth.Usage.OutputTokens)
+}
+
 func TestResponsesToAnthropic_ToolUse(t *testing.T) {
 	resp := &ResponsesResponse{
 		ID:     "resp_456",
@@ -343,6 +392,36 @@ func TestStreamingTextOnly(t *testing.T) {
 	assert.Equal(t, "message_stop", events[1].Type)
 }
 
+func TestStreamingCachedTokensUseAnthropicInputSemantics(t *testing.T) {
+	state := NewResponsesEventToAnthropicState()
+	ResponsesEventToAnthropicEvents(&ResponsesStreamEvent{
+		Type:     "response.created",
+		Response: &ResponsesResponse{ID: "resp_cached_stream", Model: "gpt-5.2"},
+	}, state)
+
+	events := ResponsesEventToAnthropicEvents(&ResponsesStreamEvent{
+		Type: "response.completed",
+		Response: &ResponsesResponse{
+			Status: "completed",
+			Usage: &ResponsesUsage{
+				InputTokens:  54006,
+				OutputTokens: 123,
+				TotalTokens:  54129,
+				InputTokensDetails: &ResponsesInputTokensDetails{
+					CachedTokens: 50688,
+				},
+			},
+		},
+	}, state)
+
+	require.Len(t, events, 2)
+	assert.Equal(t, "message_delta", events[0].Type)
+	assert.Equal(t, 3318, events[0].Usage.InputTokens)
+	assert.Equal(t, 50688, events[0].Usage.CacheReadInputTokens)
+	assert.Equal(t, 123, events[0].Usage.OutputTokens)
+	assert.Equal(t, "message_stop", events[1].Type)
+}
+
 func TestStreamingToolCall(t *testing.T) {
 	state := NewResponsesEventToAnthropicState()
 

backend/internal/pkg/apicompat/responses_to_anthropic.go

@@ -84,18 +84,34 @@ func ResponsesToAnthropic(resp *ResponsesResponse, model string) *AnthropicRespo
 	out.StopReason = responsesStatusToAnthropicStopReason(resp.Status, resp.IncompleteDetails, blocks)
 
 	if resp.Usage != nil {
-		out.Usage = AnthropicUsage{
+		out.Usage = anthropicUsageFromResponsesUsage(resp.Usage)
-			InputTokens:  resp.Usage.InputTokens,
-			OutputTokens: resp.Usage.OutputTokens,
-		}
-		if resp.Usage.InputTokensDetails != nil {
-			out.Usage.CacheReadInputTokens = resp.Usage.InputTokensDetails.CachedTokens
-		}
 	}
 
 	return out
 }
 
+func anthropicUsageFromResponsesUsage(usage *ResponsesUsage) AnthropicUsage {
+	if usage == nil {
+		return AnthropicUsage{}
+	}
+
+	cachedTokens := 0
+	if usage.InputTokensDetails != nil {
+		cachedTokens = usage.InputTokensDetails.CachedTokens
+	}
+
+	inputTokens := usage.InputTokens - cachedTokens
+	if inputTokens < 0 {
+		inputTokens = 0
+	}
+
+	return AnthropicUsage{
+		InputTokens:          inputTokens,
+		OutputTokens:         usage.OutputTokens,
+		CacheReadInputTokens: cachedTokens,
+	}
+}
+
 func responsesStatusToAnthropicStopReason(status string, details *ResponsesIncompleteDetails, blocks []AnthropicContentBlock) string {
 	switch status {
 	case "incomplete":
@@ -466,11 +482,10 @@ func resToAnthHandleCompleted(evt *ResponsesStreamEvent, state *ResponsesEventTo
 	stopReason := "end_turn"
 	if evt.Response != nil {
 		if evt.Response.Usage != nil {
-			state.InputTokens = evt.Response.Usage.InputTokens
+			usage := anthropicUsageFromResponsesUsage(evt.Response.Usage)
-			state.OutputTokens = evt.Response.Usage.OutputTokens
+			state.InputTokens = usage.InputTokens
-			if evt.Response.Usage.InputTokensDetails != nil {
+			state.OutputTokens = usage.OutputTokens
-				state.CacheReadInputTokens = evt.Response.Usage.InputTokensDetails.CachedTokens
+			state.CacheReadInputTokens = usage.CacheReadInputTokens
-			}
 		}
 		switch evt.Response.Status {
 		case "incomplete":