// changelog

v1.21.2

Gemini thinking models no longer truncate the review JSON or under-report cost, and the eval corpus grows to 129 fixtures.

Released October 31, 2026

Fixed

Gemini thinking models no longer truncate the review JSON or under-report cost. Thought tokens share MaxOutputTokens with the visible answer, so the 4096 default was often spent on thinking alone (unexpected end of JSON input / empty content). gemini-3.8-flash, gemini-3.5-flash and gemini-3.1-pro-preview now default to a 16384 ceiling, thought tokens are counted in the reported output tokens so the verbose footer, cost and eval figures match the bill, and a free-form completion cut off at MAX_TOKENS is an explicit error instead of silently truncated text. When a thinking model spends the whole output budget and returns no visible text, the request is retried once at twice the ceiling (up to the model maximum) and both calls’ usage is summed.

Eval corpus grown from 40 to 129 fixtures (96 buggy carrying 99 planted defects, 33 clean controls, 26 held-out) across Go, TypeScript, JavaScript, Python, Java and PHP, adding cross-file, large-refactor and framework-idiomatic cases.