Source: Gemini API Docs — Thinking

Google's reference for Gemini thinking: thinking levels per model, thought summaries, thought signatures and billing.

Contents
  • What: official API documentation page
  • Publisher: Google (Gemini API / AI Studio)
  • Date: accessed 2026-11-14 (examples use Api-Revision: 2026-05-20)
  • Original: Thinking (formerly ai.google.dev/gemini-api/docs/thinking)

Summary

Lists which Gemini models think, the thinking_level control and per-model defaults, how to request thought summaries, and how thought signatures carry reasoning state across turns.

Key claims

  • Thinking is on by default for all listed Gemini models (Gemini 4 family, 3.x Flash/Pro, 2.5 Pro/Flash), except gemini-2.5-flash-lite, where it is off.
  • thinking_level: minimal, low, medium, high, plus xhigh on gemini-4-hydrogen only. minimal is unsupported on Gemini 4 and 3.8 Flash. Defaults vary: high on hydrogen and 3.x Pro, medium on carbon/argon and recent Flash, minimal on Flash-Lite.
  • Thought summaries are off by default (thinking_summaries: "auto" to enable); they stream as thought_summary deltas.
  • Thought signatures are encrypted reasoning state on thought steps. In stateful mode the server manages them; in stateless mode resend every thought block unmodified, including built-in tool signatures.
  • Thinking tokens are billed as output (total_thought_tokens) for the full reasoning, not just the summary. max_output_tokens includes thinking; hitting it mid-reasoning gives an incomplete response that is still billed.
  • Guidance: minimal thinking for retrieval/classification, default for moderate tasks, maximum for hard coding, math and planning.

Feeds

Reasoning models · Extended thinking

Suggest changes →