Gemini 4 is still in refinement before an early output
Koray Kavukcuoglu, DeepMind's new chief, told The Information's summit that Gemini 4 is in post-training refinement and that Google aims to release it much earlier than the end of the year. Encouraged by the results, the team wants an early version out as soon as possible, followed by fast iterations. Safety testing continues and engineers already use the model internally in Antigravity; no date was given.
Artificial Intelligence··Midday
Kavukcuoglu said Gemini 4 is in refinement
The Verge wrote that DeepMind's new chief, Koray Kavukcuoglu, told The Information that Gemini 4 is in its refinement stage. Google aims to release it much earlier than the end of the year. The stated intention is to put out an early post-training output as soon as possible. He said the team sees the results. The Decoder reported the same appearance: Gemini 4 is in an early phase of post-training, the stage where a base model is refined so it behaves reliably. Kavukcuoglu said he hopes to release it much earlier than the end of the year.[1], [2]
Kavukcuoglu spoke for the first time as division head
It was Kavukcuoglu's first press appearance as head of the DeepMind division. A former technology chief, he took over when Demis Hassabis stepped down in August; Hassabis said at the time that Gemini 4 was making progress. Anthropic and OpenAI updated their flagship models in recent weeks. Kavukcuoglu answered the suggestion that Google has fallen behind by saying he has the utmost trust in the team and is certain Google will always be at the frontier.[1], [2], [3]
Gemini 3.5 Pro was skipped and Gemini 4 is next
Google's last major flagship releases were Gemini 3 Pro in November 2025 and Gemini 3.1 Pro in February 2026, followed by Flash updates, most recently Gemini 3.8 Flash. Gemini 3.5 Pro, teased over the summer, was never released. Kavukcuoglu said Google took a step back to focus on its Flash models and that the current focus is Gemini 4. The model is going through safety testing and guardrails, and engineers already use it internally in the Antigravity coding tool. On artificial general intelligence, he said the real question is whether we can build intelligent agents we can trust. Rival models are said to beat Gemini 3, but no benchmark was named for it.[1], [2], [3]
