Gemini Thought Summaries: Audit Deep Think Reasoning

On June 22, 2026 Google switched on Deep Think for Gemini 2.5 Pro, and the reasoning leaderboard moved overnight: 82.4% on GPQA Diamond, 89.8% on MMLU-Pro, and a record 94.1% on HumanEval+. Deep Think is an extended-reasoning configuration that explores several lines of attack in parallel before it commits to an answer, which is exactly why it does so well on hard science, math, and multi-step coding.

The benchmark headline is fun, but the part that actually changes how you build is quieter: when a thinking model reasons, the Gemini API can hand you a thought summary — a readable trace of how the model got there — plus a thought signature that lets you carry that reasoning across turns. Most teams never turn either of these on, so they debug reasoning failures blind and pay for thinking tokens they never inspect.