Gemini 4 is entering "early post-training," and Google wants to release it "much earlier" than the end of the year. The news came from Koray Kavukcuoglu, Google's new DeepMind chief, during his first public appearance on September 23, 2026. In an interview at The Information's AI Agenda summit, he said the team is seeing "exciting" results, wants to ship an early post-training build "as soon as possible," and plans to "continue the fast-paced iterations" after launch.
What he actually said
Kavukcuoglu took over as senior vice president of Google DeepMind in August after serving as the company's CTO for years — this was his first media appearance in the new role. On Gemini 4, he gave three key signals: the model is in early post-training, meaning the base model is largely locked and the work now is refinement; the launch target is much earlier than year-end, not the end-of-year window many had guessed; and rapid iteration will continue after release.
He also laid out Google's research focus: instead of arguing over whether the industry has reached the AGI label, energy should go into building "trustworthy agents." Google engineers are already running Gemini 4 on the Antigravity AI coding tool, letting the model surface issues in real engineering workflows ahead of launch. Alphabet CEO Sundar Pichai said it bluntly on the Q2 earnings call: coding and agentic coding are the areas Google most needs to fix.
Context: Google's flagship gap year
To weigh this announcement, look at Google's past year. Its last flagship was the Gemini 3 series in November 2025; since then, OpenAI's GPT-6 series and Anthropic's Mythos series have shipped. Pichai previewed Gemini 3.5 Pro at the May I/O conference for a June release — it still hasn't shipped publicly and remains in partner testing.
Kavukcuoglu's explanation: Google "took a little bit of a step back" to prioritize the faster, lighter Flash models. On July 21, Google kicked off its "most ambitious pre-training run yet" — Gemini 4. Going from pretraining start to post-training in about two months is a clear acceleration for Google.
How big is the gap
Artificial Analysis' Intelligence Index v4.3.2 puts Google's strongest current model, Gemini 3.6 Flash, at 34.0 — against 57.6 for Anthropic's Claude Opus 5.5, 52.7 for OpenAI's GPT-6 Astra, and 47.5 for GPT-6 Sol. A gap of 24 points is not small: it means Google is currently competing below the frontier line. (We previously covered the launch and price cut of Claude Opus 5.5, with a fuller price-and-capability comparison.)
What an early launch means
Reaching post-training means Gemini 4's base model is essentially locked; what's left is tuning, alignment, and safety testing — the final stretch before release. Google is now putting "speed" front and center: Pichai has pushed for a cadence approaching monthly releases, and Kavukcuoglu promised fast iteration to continue after launch. For developers, the most immediate signal is that Google's flagship supply will likely resume before year-end — the option missing from API lineups for a year is coming back.
An announcement is not a delivery, of course. The Gemini 3.5 Pro no-show is still fresh, and whether Kavukcuoglu's "much earlier than year-end" holds will depend on how smoothly post-training and safety testing go. But at least one thing is concrete: Google has put the chase on a public timeline.
