Fix Gemini/Vertex estimated cost excluding thinking tokens

packages/proxy/src/providers/google-converter.ts

@@ -147,7 +147,14 @@ const geminiUsageToOpenAIUsage = (
 
   const usage: OpenAICompletionUsage = {
     prompt_tokens: usageMetadata.promptTokenCount || 0,
-    completion_tokens: usageMetadata.candidatesTokenCount || 0,
+    // Gemini reports visible output (candidatesTokenCount) and thinking
+    // (thoughtsTokenCount) separately, and totalTokenCount is their sum. Follow
+    // the OpenAI convention (also used by the Anthropic converter and Lingua)
+    // where completion_tokens includes reasoning, so downstream cost estimation
+    // prices thinking tokens as output. reasoning_tokens below stays the subset
+    // breakdown.
+    completion_tokens:
+      (usageMetadata.candidatesTokenCount || 0) + (thoughtsTokenCount || 0),
     total_tokens: usageMetadata.totalTokenCount || 0,
   };
 

packages/proxy/src/providers/google.test.ts

@@ -1412,6 +1412,37 @@ describe("googleCompletionToOpenAICompletion", () => {
     expect(result.choices[0].finish_reason).toBe("stop");
   });
 
+  it("should include thinking tokens in completion_tokens for cost estimation", () => {
+    const geminiResponse: GenerateContentResponse = {
+      candidates: [
+        {
+          content: {
+            role: "model",
+            parts: [{ text: "The answer is 42." }],
+          },
+          finishReason: "STOP",
+        },
+      ],
+      usageMetadata: {
+        promptTokenCount: 100,
+        candidatesTokenCount: 50,
+        thoughtsTokenCount: 25,
+        totalTokenCount: 175,
+      },
+    };
+
+    const result = googleCompletionToOpenAICompletion(
+      "gemini-2.5-flash",
+      geminiResponse,
+    );
+
+    // completion_tokens must include thinking tokens (50 + 25) so downstream
+    // cost estimation prices reasoning as output. reasoning_tokens is the subset.
+    expect(result.usage?.completion_tokens).toBe(75);
+    expect(result.usage?.completion_tokens_details?.reasoning_tokens).toBe(25);
+    expect(result.usage?.prompt_tokens).toBe(100);
+  });
+
   it("should map safety finish reasons to content_filter", () => {
     const geminiResponse: GenerateContentResponse = {
       candidates: [
@@ -1755,6 +1786,38 @@ describe("googleEventToOpenAIChatEvent", () => {
     });
     expect(result.event!.choices[0].finish_reason).toBe("tool_calls");
   });
+
+  it("should include thinking tokens in streaming completion_tokens", () => {
+    // eslint-disable-next-line @typescript-eslint/no-explicit-any
+    const geminiEvent: any = {
+      candidates: [
+        {
+          content: {
+            role: "model",
+            parts: [{ text: "The answer is 42." }],
+          },
+          finishReason: "STOP",
+        },
+      ],
+      usageMetadata: {
+        promptTokenCount: 100,
+        candidatesTokenCount: 50,
+        thoughtsTokenCount: 25,
+        totalTokenCount: 175,
+      },
+    };
+
+    const result = googleEventToOpenAIChatEvent(
+      "gemini-2.5-flash",
+      geminiEvent,
+    );
+
+    // eslint-disable-next-line @typescript-eslint/no-explicit-any
+    const usage = (result.event as any)?.usage;
+    expect(usage?.completion_tokens).toBe(75);
+    expect(usage?.completion_tokens_details?.reasoning_tokens).toBe(25);
+    expect(usage?.prompt_tokens).toBe(100);
+  });
 });
 
 describe("geminiParamsToOpenAITools", () => {