Google’s next flagship AI model has entered post-training, and the company says it will ship “much earlier” than the end of 2026.
Koray Kavukcuoglu, who took over Google DeepMind in August after co-founder Demis Hassabis stepped down, made the remarks at The Information’s AI Agenda Live Summit in his first public appearance since the promotion. Post-training is the phase where a finished base model is refined with human feedback, reinforcement learning and specialised data to shape reasoning, safety and reliability — the last major stage before a public release.
“Our intention is to, like — as soon as possible — release an early post-training output because we see the results and we are excited,” Kavukcuoglu told The Information. He added that Google plans to “continue the fast-paced iterations” from there.
The strategy: ship early, iterate in public
The most revealing part of the announcement is not the timeline but the delivery model. Google plans to deploy Gemini 4 in stages — an early version first, refined by real-world feedback — rather than waiting for a polished launch. That is a departure from the polished, single-event flagship releases that defined Gemini 3.
Internal testing is already underway: Gemini 4 is powering Antigravity, Google’s agentic development environment, as part of the safety testing and guardrail work that post-training involves.
Asked directly whether Google has fallen behind, Kavukcuoglu said: “I have the utmost trust in the team” and “in my mind, it’s a certainty that we are always gonna be at the frontier.”
Why Google is in a hurry
The urgency is not hard to read. Google has not released a new flagship model since the Gemini 3 series in November 2025, with Gemini 3.1 Pro following in February 2026. Since then, its rivals have kept swinging: OpenAI launched GPT-6 Astra on September 3 and followed with the cheaper GPT-6 Sol and Luna on September 22, while Anthropic shipped Claude Fable 5.1 on September 1 and Claude Opus 5.5 on September 22.
Google’s own June promised a Gemini 3.5 Pro update that never arrived. Pichai announced it at I/O in May; by summer the model had been quietly shelved. Kavukcuoglu admitted Google “took a little bit of a step back” to focus on its faster, cheaper Flash models instead. The market has noticed the gap — Alphabet shares fell nearly 5 percent over two sessions this week as Meta’s Muse agent climbed the US app charts during Meta Connect.
There is also an unusual leak circulating: developers testing an unnamed entry on LMArena labelled “gemini-3.8-flash” — a designation that already belongs to a real, shipping model — believe it is actually an early Gemini 4 Pro checkpoint, with a circulating benchmark table ranking it first on agentic coding and knowledge work. Google has not confirmed the entry is Gemini 4, and leaked benchmark tables are unverified by nature, but the timing lines up with a model in final refinement.
What it means
If Gemini 4 does arrive well before year-end, it will land in the middle of the densest release window the frontier has seen — and Google’s staged-release approach means it will be judged on iteration speed, not launch-day polish. The company’s own framing has shifted too: asked about AGI, Kavukcuoglu said the conversation “is more about are we able to build intelligent agents that we can trust.”
Google has been here before with its robotics push, where DeepMind’s whole-body control models run alongside — not ahead of — its frontier language releases. And the delay-and-recover pattern is familiar: Gemini 3.5 Pro’s slide past its July window signalled the same race intensification now forcing Google’s hand.
Whether “as soon as possible” means November or January is the part Google is not saying.
