Gemini Text Agent: Thinking Level for Gemini 3.x vs Sampling Parameters

For a PoC we’re using the Gemini Text Agent with the Flash model, which resolves to Gemini 3.8 Flash. The action (still) exposes the sampling parameters: Temperature, Top P, and Top K.

According to Google’s documentation, these sampling parameters are deprecated for Gemini 3.x and ignored by the backend, and determinism is meant to be controlled through thinking_level (LOW / MEDIUM / HIGH) and a response schema instead.

Question: Has anyone tried the Zenphi HTTP action calling the Gemini API directly with thinking_level set? (while we wait for Zenphi to update the Gemini actions)

Use case: legal validation of extracted data against supplied source documents, where run-to-run consistency matters.