Google DeepMind's new chief speaks: Gemini 4 is in post-training and will land well before year end

Start with who is talking, because it determines how much weight the story carries. Koray Kavukcuoglu was Google DeepMind's CTO, and took over day-to-day operations of the division in August 2026 after Demis Hassabis stepped down as CEO, with Hassabis becoming chair of the unit and Alphabet's chief scientist. According to The Decoder's reporting, Kavukcuoglu does not carry the CEO title, and Google is consolidating Gemini development in the Bay Area. His appearance was at The Information's AI Agenda Live Summit, his first public media interview since taking over. So these are not a product manager's teasers, they are the new head's first public statements.

The second thread is what the phrase post-training actually means. It is the stretch of work that turns a base model into a reliably behaving product, and per 9to5Google that includes safety testing and implementing guardrails. In other words, the capability base is settled and what remains is behavioural engineering. That explains his widely quoted line: "Our intention is to, like - as soon as possible - to release an early post-training output because we see the results and we are excited... [to] continue the fast-paced iterations." There is no date in that sentence, only a judgement: the results are good enough to be worth shipping early. 9to5Google adds that internal testing is already using Gemini 4 to power Antigravity, which is a checkable usage rather than a press-line.

The third thread is the gap in the timeline. Google's last flagship was Gemini 3 in November 2025, followed by Gemini 3.1 Pro in February 2026, after which the releases were mostly smaller Flash models, most recently Gemini 3.8 Flash on September 2. In between there was an episode: per 9to5Google and The Decoder, CEO Sundar Pichai announced at the May I/O that Gemini 3.5 Pro would arrive in June, and the model never shipped. The Decoder cites the Financial Times saying the project missed three deadlines. Kavukcuoglu's framing is that Google took "a little bit of a step back" to maximise the speed of learning at the time, and he did not say whether Gemini 3.5 Pro has been formally cancelled.

The fourth thread is competitive position. Google's lag is not a feeling but has concrete reference points: Anthropic updated its Mythos/Fable family this month and shipped Opus 5.5, while OpenAI has been rolling out GPT-6 models since early September. Asked whether he worries Google has fallen behind, Kavukcuoglu said he has "the utmost trust in the team," that he is "a hundred percent confident," and that "in my mind, it's a certainty that we are always gonna be at the frontier." He also raised an angle that is easy to miss: the model work feeds Google's chip business, since the company began selling TPUs directly to customers this year, and new models show hardware engineers how to design the next two or three generations of chips.

The fifth thread requires separating facts from third-party analysis. The fact is that Kavukcuoglu said plainly: "The conversation of [whether] we achieved AGI or not is not the right conversation," adding that the right one is "are we able to build intelligent agents that we can trust?" The Decoder's reading is that a lab long defined by the pursuit of artificial general intelligence now treats progress as a product question, turning from a research organisation into Google's Gemini factory. The same piece catalogues the departures: Noam Shazeer joined OpenAI in June, AlphaFold Nobel laureate John Jumper then left for Anthropic, and Jeff Dean founded Discovery Loop with Oriol Vinyals, Quoc Le and Sanjay Ghemawat; it cites Fortune data showing DeepMind's ratio of new hires to departures fell from about 12 to 1 in Q2 2023 to roughly 2 to 1 in Q3 2026. On whether Hassabis' exit was voluntary, The Decoder cites the newsletter Pathfounders and notes Google has not commented. That item is unconfirmed reporting and has to be labelled as such.