Gemini 4 Enters Post-Training, and the Race Tightens

Google's next flagship model is closer than the calendar suggested. Speaking at The Information's AI Agenda Live summit on September 23, Koray Kavukcuoglu, the senior vice president who now runs Google DeepMind, said Gemini 4 has entered post-training and that he expects it to arrive "much earlier" than the end of the year.

Post-training is the stage where a raw, pre-trained model is shaped into something usable. The base model has already absorbed its training data. What remains is the tuning that teaches it to follow instructions, hold a conversation, use tools, and refuse the things it should refuse. It is the last major phase before a model is ready to face the public, and reaching it is a meaningful signal about timing.

Kavukcuoglu was careful not to promise a date. He described Gemini 4 as being in the early part of post-training, which is a window that can run anywhere from a few weeks to a couple of months depending on how the evaluations go. Google has not published benchmarks, pricing, or a feature list, and until it does, the only firm fact is that the model exists and is being polished.

A three-way race with almost no gap

The timing is what makes the remark notable. Gemini 4 is landing into the tightest stretch of competition the field has seen. In the space of a single week, Anthropic shipped Claude Opus 5.5 and OpenAI answered with sharp price cuts, and xAI released Grok 4.7. Each of those launches leapfrogged the last on some benchmark or another. Google has spent much of the year a step behind on release cadence, even as its models scored well once they arrived.

Kavukcuoglu's message, in effect, was that the gap has closed. A model in post-training now is a model that can respond to whatever its rivals shipped this month rather than answering last quarter's releases. That is a different posture for Google, which has often looked like it was reacting rather than setting the pace.

What to watch

The interesting questions are the ones Google has not answered yet. Whether Gemini 4 leads on the reasoning and coding evaluations that buyers now fixate on. Whether it ships across the full Gemini lineup at once or starts with a single tier. And whether Google prices it to defend market share after OpenAI cut its own rates to the bone days earlier.

For now, the takeaway is narrow but real. The company that many counted as the slow mover in this cycle says its next frontier model is nearly done. If Kavukcuoglu's "much earlier" holds, the year is going to end with all four major labs shipping within weeks of each other, and buyers, for once, holding the leverage.