[{"h":"CONFIRMED: Google moves Gemini 4 out of rumor territory","p":"Since July, Gemini 4 lived only on indirect leads: an ambitious pre-training announcement, a checkpoint disguised as gemini-3.8-flash running on LMArena, and a leaked, unaudited benchmark table. That changed on September 23. At The Information's AI Agenda Live Summit, Koray Kavukcuoglu, who took over Google DeepMind's technical leadership after Demis Hassabis moved to chair of Alphabet, said on the record that Gemini 4 is in post-training, the final stage before any public release, and that the intent is to ship an initial version as soon as possible, iterating quickly on real user feedback. It is the first time anyone at Google has spoken about the model with a name, title and timeframe since the pre-training announcement on July 21."},{"h":"What the statement changes","p":"Kavukcuoglu was direct about timing: the model should arrive well before the end of 2026, pushing the market from a speculated November-December window to a matter of weeks. He also described the release shape as staged, an initial version followed by rapid updates, rather than a single finished package. That echoes Google's own recent pattern: four successive Flash models shipped since May, while Gemini 3.5 Pro was scrapped after being announced because internal candidates did not beat the company's own Flash line. When the division's technical chief says publicly the company intends to keep that fast cadence, the practical read is that Gemini 4 will likely arrive in slices, not a single event."},{"h":"What is still rumor","p":"Kavukcuoglu's statement confirms the model exists, is advanced and has a timeframe, but confirms nothing about the performance claims already circulating. The checkpoint labeled gemini-3.8-flash that surfaced on LMArena in mid-September, with convergent tester reports and a leaked table putting the alleged Gemini 4 Pro ahead of GPT-6 Astra and Claude Fable 5.1 on DeepSWE, GDPval-AA, Terminal-bench and OSWorld, remains without any confirmation, official name, model card or pricing page from Google. The same goes for the leaked price of $2.25 and $11.25 per million tokens and the 2-million-token context window: those are numbers from an unaudited table, not official documentation."},{"h":"Why separating the two matters","p":"Mixing the calendar confirmation with the benchmark leak would be the classic error of treating two very different strengths of evidence as one. One is a public statement from a named executive, at a press event, about a real development phase. The other is a swapped label on a comparison arena and a spreadsheet with unclear provenance. The first upgrades this article from STRONG SIGNAL to CONFIRMED regarding the model's existence and timeline. The second stays exactly where it was: a strong signal about performance, not a fact."},{"h":"MaxAssistant's read","p":"For the first time in this story, Google chose to talk about Gemini 4 instead of letting the community fill the vacuum. That is a good sign the model is real and close, but it does not settle whether it will actually beat Astra and Fable 5.1 on the numbers already circulating. Anyone planning a production stack should treat September as the month the launch moved from pure hypothesis to the calendar, and hold judgment on performance until Google publishes a real model card and price."},{"h":"Sources","p":"The Information, AI Agenda Live Summit, remarks by Koray Kavukcuoglu (September 23, 2026)
TechBriefly, Leaked Gemini 4 Pro benchmarks show it beating GPT-6 Astra and Claude
Dataconomy, Google May Be Secretly Testing Gemini 4 Pro"}]
Releases & Models
CONFIRMED: Google admits Gemini 4 is already in post-training and promises a 2026 launch
Koray Kavukcuoglu, Google DeepMind's new head, publicly confirmed that Gemini 4 is in post-training and should arrive well before the end of 2026, in a staged rollout. It is the first official statement about the model since July. The gemini-3.8-flash checkpoint that leaked disguised on LMArena remains unconfirmed by Google.