Google's next model arrives.

Google released Gemini 3 yesterday, its first major model update since December 2024.

The system was trained on Google's TPU hardware, which is the native chip platform they use for large-scale model training. Early tests show a significant increase (2X SOTA) on the ARC reasoning benchmark, a test often used to measure AGI progress.

The release follows a period of rapid updates across the industry. GPT-5.1, Grok 4.1, and Kimi 2 all shipped within weeks, and Google is re-entering the cycle after spending much of the year focused on partnerships and infrastructure. They have also expanded their partnerships, notably with Anthropic, and continued to make TPUs available to other AI companies.

New models take time to settle, but it's clear that companies are continuing to invest billions in training hardware, model reliability, and each release provides developers with more capabilities to work with.