Gemini-3.1-pro-preview: thoughtsTokenCount missing with thinking_level=LOW

Gemini API
1

Hi,

I’m seeing inconsistent usage metadata for gemini-3.1-pro-preview on Vertex AI.

When using:

  • thinking_level=HIGH → thoughtsTokenCount is present
  • thinking_level=MEDIUM → thoughtsTokenCount is present
  • thinking_level=LOW → thoughtsTokenCount is missing

This happens both in:

  • batch inference
  • online inference

This worked as expected in gemini-3-pro-preview (thoughtsTokenCount always present).

Is this expected behavior for LOW, or a bug?
This makes it difficult to accurately estimate billed tokens from usageMetadata.

Thanks!