Gemini 2.5 Calls Reroute to Gemini 3 on Dec. 1 and Can Break Apps

Google will automatically reroute all Gemini Developer API requests to `gemini-2.5-pro` to `gemini-3.8-flash` and requests to `gemini-2.5-flash` to `gemini-3.5-flash-lite` on Dec. 1, 2026, according to the Gemini API changelog published Nov. 2, 2026, with billing following the new models' rates. Google listed four breaking changes for both deprecated models: function calling and `thought_signature` validation, `thinking_level` replacing `thinking_budget`, removal of sampling parameters, and removal of prefilled model turns; Firebase's Nov. 2 release notes said some settings will return 400 errors after the switch. Google advised developers to change the model string before Dec. 1, test function calling and `thought_signature` handling, swap `thinking_budget` for `thinking_level`, and strip sampling parameters and prefilled turns.

Gemini 2.5 Calls Reroute to Gemini 3 on Dec. 1 and Can Break Apps Change the model string and drop thinking budget and sampling parameters before Dec. 1, or Gemini 2.5 requests get rerouted to Gemini 3 models. Any app that still calls gemini-2.5-pro or gemini-2.5-flash through the Gemini Developer API will have those requests rerouted to Gemini 3 models on Dec. 1, with different billing and request rules. Google announced the change in its Nov. 2 Gemini API release notes Gemini API changelog, Nov. 2, 2026 https://aistudio.google.com/docs/changelog . What Google Said Both models are deprecated and no longer selectable in Google AI Studio. “Effective December 1, 2026, all API requests to gemini-2.5-pro will automatically route to” gemini-3.8-flash , Google said, and billing follows Gemini 3.8 Flash rates. Requests to gemini-2.5-flash route to gemini-3.5-flash-lite at Flash-Lite rates. Google listed four breaking changes for both models: - Function calling and thought signature validation - thinking level replaces thinking budget - Sampling parameters are removed - Prefilled model turns are removed Google’s migration advice: change the model string before Dec. 1, test function calling and thought signature handling, swap thinking budget for thinking level , and strip sampling parameters and prefilled turns. It linked a migration guide for each target model. Firebase repeated the notice in its Nov. 2 release notes, and said some settings will return 400 errors after the switch. The Take A silent reroute is a worse deal than a shutdown. A shutdown errors loudly on day one. A reroute keeps returning 200s from a different model with different pricing, and requests that use a setting the new model rejects fail with a 400. Firebase AI Logic apps that name a 2.5 model in client code are the likeliest to get caught. A shipped mobile build keeps its hardcoded model string until users update. Server code and Genkit flows change in one deploy. Grep every repo for gemini-2.5 and for thinking budget , temperature , top p and top k sent to those models. Ship the new model string in a test build now, not on Nov. 30. For mobile apps, the update lag means the work should start this week.