Google DeepMind’s new chief, Koray Kavukcuoglu, wants Gemini 4 out well before the end of the year. His priority is to deliver an early version and learn from rapid follow-up releases, while the company continues safety testing and develops safeguards.
The plan also reflects a broader change at DeepMind: a lab long associated with artificial general intelligence is now centered on building Google’s Gemini products.
Gemini 4 is in post-training
At his first media appearance as head of Google DeepMind, at The Information’s AI Agenda Live Summit, Kavukcuoglu said Gemini 4 was in an early stage of post-training. This is where a base model is refined to behave more reliably.
He said Google hopes to release an early post-training output “much earlier” than the end of the year. Kavukcuoglu said the team is encouraged by the results, and that fast iterations would follow. Google is still working on safeguards and safety testing, while engineers are already using Gemini 4 internally with the coding tool Antigravity.
Kavukcuoglu, previously the company’s CTO, took over in August after Demis Hassabis stepped down as DeepMind CEO. Hassabis now chairs the unit and serves as Alphabet’s chief scientist.
A flagship release after a missed launch
Gemini 4 would be Google’s first new flagship model since Gemini 3 late last year. Since then, Google has shipped the Gemini 3.1 update in February and several smaller, cheaper Flash models.
The company had also announced Gemini 3.5 Pro at I/O in May for a June release, but it never arrived. The Financial Times reported that the project missed three deadlines. Kavukcuoglu said Google took “a little bit of a step back,” explaining that the priority was to maximize the speed of learning. He did not say whether Gemini 3.5 Pro had been officially canceled.
That history gives the Gemini 4 timeline added weight. A quicker early release could give Google a chance to refine its model through repeated work, but the source does not establish how the final product will compare with rivals.
In the meantime, Anthropic released its Mythos/Fable family in the spring and updated it, along with the Opus line, this month. OpenAI has been rolling out GPT-6 models since early September. Kavukcuoglu said he was “a hundred percent confident” in his “insanely strong team,” adding: “In my mind, it's a certainty that we are always gonna be at the frontier.”
From AGI ambitions to trustworthy agents
The change in leadership is also visible in how Kavukcuoglu frames the work. Asked about artificial general intelligence, he said: “The conversation of [whether] we achieved AGI or not is not the right conversation.” Instead, he said the question is whether the company can build intelligent agents people can trust.
That emphasis differs from Hassabis’s farewell memo, in which he wrote: “I've been working towards AGI my whole life and now, like many of you, I feel it is close at hand.” Hassabis co-founded Deepmind in 2010, and Google acquired the lab four years later.
DeepMind had enjoyed special status within Google for years. ChatGPT changed the company’s priorities: during its first “Code Red,” Google combined Google Brain and DeepMind into Google DeepMind. The Gemini models got their name from that combined organization.
The August restructuring pushed the shift further. Kavukcuoglu now runs day-to-day operations without the CEO title, and all Gemini development is moving to the Bay Area. The former research lab’s work is increasingly organized around delivering Google’s AI products.
Models also shape Google’s chip plans
Gemini development matters to Google beyond the models themselves. The company started selling its TPUs directly to customers this year, bringing it into closer competition with Nvidia. Kavukcuoglu said new models help hardware engineers design the “next two or three generations” of chips.
Google has caught up before: Gemini 3 was such a leap that OpenAI declared a “Code Red” and only pulled even again with GPT-5.2. Whether Gemini 4 can produce a similar shift remains an open question.
For now, the clearest signal is the new chief’s focus: release an early model, iterate quickly, and judge progress by whether the resulting agents can be trusted. That puts product delivery at the center of DeepMind’s next chapter, even as the debate over AGI and its risks continues.