Google Scraps Gemini 3.5 Pro's Architecture and Rebuilds It From Scratch, Targeting July 17

My take: Scrapping a frontier model and rebuilding from scratch is expensive — in time, money, and engineering capital. Google DeepMind doing this with Gemini 3.5 Pro suggests the team found structural flaws that could not be patched: the original architecture reportedly failed on recursive tool calls and SVG generation.

Leaks point to a 2-million-token context window (double Gemini 2.5 Pro) and a deep reasoning mode called Deep Think. None of those figures have been officially confirmed by Google — treat them as directional signals, not signed specs. That also means when benchmarks do arrive, it will be worth waiting for independent third-party testing before taking Google's own numbers at face value.

By July 17, when the launch is expected, GPT-5.6 Sol and Fable 5 will have been in production for weeks. For anyone evaluating which model to use in their work: the extra time Google took may mean a more solid product, or it may mean they missed the moment. We'll find out on the 17th.