Gemini 4 in early post-training

Gemini 4, Google’s next flagship frontier model, is moving through early post-training with a launch targeted much earlier than the end of 2026, according to the newly appointed head of Google DeepMind, Koray Kavukcuoglu. He said the team intends to “release an early post-training output as soon as possible” and keep up the fast iteration cadence.

What’s actually new

• Post-training phase. The model has moved from the pre-training run Google confirmed on July 21, 2026, into refinement, where a base model is tuned for reliable real-world performance.
• Public timeline for the first time. Google had announced Gemini 4’s existence but never a launch window; Kavukcuoglu’s remarks are the first concrete timing signal, well before prior analyst estimates of late 2026.
• Context. Google is racing Anthropic and OpenAI on frontier models after Gemini 3.5 Pro slipped on coding performance, and Gemini 4 is its answer aimed squarely at the frontier level.

Why it matters

This is an “in development” update, not a launch: no release date, benchmarks, prices, or API model ID have been published. Declared via an on-record interview rather than a first-party announcement, so treat the timing as directional. For builders, it’s a heads up that a new Google flagship is coming soon and early access may appear quickly.

(Original source: The Information)