Google DeepMind Chief Says Gemini 4 Is in Post-Training, Aims for Release This Year

Gemini 4 has reached early post-training, and Google DeepMind chief Koray Kavukcuoglu hopes to release an early version well before the end of 2026. In an interview at The Information’s AI Agenda Live Summit, he said Google is still working on safeguards and safety testing. The model is also being used internally, though Kavukcuoglu gave no release date.

An early version before the work is done

Post-training is the phase after initial training in which a model’s behavior is refined. Kavukcuoglu said Google intends to release an early output from that phase “as soon as possible,” then iterate quickly. His year-end target is a hope, not a dated launch commitment.

An early output would not mean Google had finished its safety work. Nor did Kavukcuoglu specify when people outside the company could try Gemini 4. Those are separate questions for anyone waiting to put the model to work.

A coding tool becomes the proving ground

Google engineers are already using Gemini 4 internally in Antigravity, the company’s coding tool. That gives the team a setting for assessing how the model works inside a software task. Internal use does not establish how it will perform for outside users.

Kavukcuoglu described the aim in an earlier Google AI conversation: moving from a model that helps with coding toward an agent that can work with software tools. At the time, he said Gemini 4’s potential would have to be realized first with internal users, then external ones.

The Antigravity work gives that ambition a concrete test, but not a public verdict. The question is whether an agent can be trusted when it helps carry out a task, rather than simply suggesting an answer.

The conversation is more about are we able to build intelligent agents that we can trust?

Koray Kavukcuoglu at The Information’s AI Agenda Live Summit, as quoted by The Decoder

Trust, rather than an AGI finish line

Kavukcuoglu called the question of whether AI has achieved artificial general intelligence, or AGI, “not the right conversation.” That is a change in emphasis, not evidence that Google DeepMind has dropped the goal. In the earlier Google conversation, he said no single test could establish AGI and described trust as something built gradually.

His predecessor, Demis Hassabis, struck a different note in a farewell memo, writing that AGI felt close at hand. Hassabis stepped down as DeepMind CEO in August, and Kavukcuoglu took over day-to-day leadership. Their remarks offer different ways to describe progress: an anticipated milestone, or agents people can rely on.

The gap left by Gemini 3.5 Pro

The Gemini 4 target follows an unfulfilled plan for Google’s previous model line. CEO Sundar Pichai announced Gemini 3.5 Pro in May for a June release, but it had not arrived by September. Kavukcuoglu said Google had taken “a little bit of a step back” from that model; he did not say whether it was canceled.

Google released Gemini 3.1 in February and several smaller Flash models after Gemini 3 arrived late last year. Kavukcuoglu now says the team’s focus is Gemini 4. The earlier missed target makes the distinction between a hoped-for early output and a usable public release especially important.