Spotted here for gemini-2.5-pro-preview-05-06: https://gist.github.com/simonw/87a59e7f5c12274d65e2ac053b0eacdb#token-usage

264 input, 104 output, {"promptTokensDetails": [{"modality": "TEXT", "tokenCount": 6}, {"modality": "IMAGE", "tokenCount": 258}], "thoughtsTokenCount": 989}

And then Aider wrote about the same problem: https://aider.chat/2025/05/07/gemini-cost.html

An investigation determined the primary cause was that the litellm package (used by aider for LLM API connections) was not properly including reasoning tokens in the token counts it reported.

Here's where LiteLLM fixed it: BerriAI/litellm@a7db0df

This note is very important:

• Gemini-2.5-flash - support reasoning cost calc + return reasoning content BerriAI/litellm#10141 (comment)

Note that the Gemini API returns different usage metadata than Vertex AI. With the Gemini API, candidatesTokenCount includes thinking tokens, but on Vertex AI, candidatesTokenCount does not include thinking tokens.

But that doesn't fit what I'm seeing here, because I did NOT use Vertex but still got that response where thoughtsTokenCount did not get included in the candidatesTokenCount

For reference, here's our current code for that: