Google DeepMind chief Koray Kavukcuoglu said at The Information’s AI Agenda Live Summit on 23 September that Gemini 4 had entered early post-training and could launch “much earlier” than the end of the year, Benzinga reported on 24 September. Google wants to release an early output from that stage as soon as possible, according to remarks reported by The Verge.
Key points
- Gemini 4 is in early post-training, the stage in which its base model is refined for more reliable performance.
- Google expects a launch well before the end of 2026, though the expectation is not a release date.
- Google’s previous flagship series arrived in November 2025; a planned Gemini 3.5 Pro update did not ship.
Gemini 4 moves into post-training
Post-training begins with a base model that has already undergone its main training run. The subsequent work adjusts how that model responds, with more reliable performance the aim Kavukcuoglu described. It is something like tuning an engine after building it: the underlying machinery is there, but its behaviour when put to use still matters. His description places Gemini 4 in that refinement stage, rather than at the point of a finished release.
Writing a reply to a routine email could become more dependable if that refinement makes responses more reliable. The activity would still depend on the finished model performing as intended, rather than merely having passed through another stage of training.
Kavukcuoglu said Google intended to release an early post-training output as soon as possible because it was encouraged by its results, The Verge reported. That describes the company’s assessment of work under way, not a measured result for a model people can use. His expectation that Gemini 4 would arrive much earlier than year-end gives a broad window for release, rather than a particular day.
In July, Google chief executive Sundar Pichai said the company had begun what he called its “most ambitious pretraining run yet” for Gemini 4, Benzinga reported. That was Pichai’s description of Google’s own run. Pre-training builds the base model that post-training then refines, so the two comments concern successive parts of the work, rather than two measurements of how well Gemini 4 performs.
Gemini 3 remains Google’s latest flagship
Google’s Gemini 3 series arrived in November 2025 and remained its latest flagship model at the time of Kavukcuoglu’s remarks, The Verge reported. OpenAI had since released GPT-6 and Anthropic its Mythos series. Those releases put the timing of Google’s next flagship in focus for companies choosing which models to build upon, even though a release timetable by itself says nothing about a model’s performance in their applications.
The Verge described both rival series as outperforming Gemini 3. For work that depends on a model answering consistently, the post-training effort is relevant precisely because Kavukcuoglu described reliability as its purpose.
Kavukcuoglu took over as its chief after Demis Hassabis stepped down from the role in August, The Verge reported. The summit appearance was Kavukcuoglu’s first with the media as the division’s leader. He also serves as Google’s chief AI architect, a role that places his account of the model’s progress within the company’s own plans.
Gemini 3.5 Pro gave way to Flash
Before Gemini 4, Google had planned a Gemini 3.5 Pro update. Pichai said in May that it would come in June, but the update did not ship. A planned release that falls away matters to companies building on these models because they can use a model that arrives, rather than one whose expected arrival has passed.
Kavukcuoglu said Google took a step back from the Gemini 3.5 Pro update to concentrate instead on Flash models, which are faster and less powerful. That choice concerned what Google put forward while work on its next flagship continued. It also accounts for why the absence of a new flagship did not mean an absence of model releases: the company had focused on a different kind of model from the Pro update it had planned.
