Gemini 4 is already in post-training: DeepMind's new chief promises a launch well before year's end

Google has finished the heaviest part of building its next flagship AI model. Speaking publicly for the first time since taking over Google DeepMind, Koray Kavukcuoglu said Wednesday that Gemini 4 has entered early post-training — the refinement phase that follows the main training run, where a finished base model is tuned for reliable, safe behavior before it ships.

The promise on timing was the headline. According to The Information, which hosted the summit where he spoke, Kavukcuoglu said Google intends to roll out an early post-training version "as soon as possible" — "much earlier" than the end of the year — and then keep iterating at a fast pace. It is the most concrete release signal Google has given for Gemini 4, and it arrives in a month when rivals have been shipping flagships almost weekly.

What Kavukcuoglu actually said#

The setting was The Information's AI Agenda Live Summit on September 23, Kavukcuoglu's first public appearance since becoming the head of DeepMind. "Our intention is to roll out an early post‑training version as soon as possible, because we've already seen promising results and are very excited," he said, according to the outlet's account of the interview.

The stage of training matters here. Pre-training — the giant, expensive run that bakes capabilities into the base model — is effectively done; CEO Sundar Pichai described the effort on Alphabet's second-quarter earnings call as the company's "most ambitious pre-training run yet," with serious training underway since July. Post-training is what turns that raw capability into something a developer can call: instruction tuning, safety testing, behavior refinement.

One caveat worth keeping in mind: no firm date was given. "Much earlier than year's end" is an aspiration, and frontier-model timelines have a long history of slipping during safety testing. But even as an aspiration, it puts Google's next flagship on the calendar — something the company had conspicuously avoided doing until now.

A new chief, a new tempo#

The messenger matters as much as the message. Kavukcuoglu took the DeepMind helm in August after Demis Hassabis stepped down as CEO to become chairman of the unit and Alphabet's chief scientist. Kavukcuoglu had been the company's chief technology officer — and his first outing suggests a real change in posture.

Where Hassabis spent years framing DeepMind's mission around artificial general intelligence, his successor reportedly called the AGI question "not the right conversation." The framing Kavukcuoglu offered instead: "The conversation is more about are we able to build intelligent agents that we can trust." Less philosophy, more product.

That posture fits what Google actually did in 2026. The flagship line went quiet: Gemini 3.5 Pro, previewed at I/O for a June launch, never shipped. Google instead, in Kavukcuoglu's telling, "took a little bit of a step back" after the 3.x series to pour effort into the lighter Flash line — new Flash variants arriving almost monthly — and, crucially, into learning how to ship faster. Pichai has said he wants Gemini on nearly a monthly release cadence, a sharp break from the lab's historically slower rhythm.

Why Google is racing the clock#

The competitive ledger explains the urgency. OpenAI launched its flagship GPT-6 Astra in early September. Anthropic unveiled Claude Fable 5.1 on September 1. Then on September 22, both labs shipped new models on the same day — GPT-6 Sol and Luna from OpenAI, Claude Opus 5.5 from Anthropic — each with steep price cuts. Google hasn't shipped a new flagship model since the Gemini 3 series late last year.

The pressure is not abstract. Alphabet shares fell nearly 5% over two sessions this week as Meta's Muse AI agent gained consumer traction, then steadied after Kavukcuoglu's remarks. And the battleground is increasingly clear: agentic coding. Both rivals are winning developer mindshare with models built for long-horizon coding work, and Pichai has singled out coding and agentic coding as the areas Google most needs to fix.

That is presumably why Gemini 4 is being dogfooded inside Antigravity, Google's agentic software-development platform, where engineers are already running the unreleased model on real coding tasks. The real test for Gemini 4 won't be a benchmark chart — it will be whether developers trust it to carry out work inside their own tools.

The talent tradeoff#

There is a second, quieter story running alongside the launch timeline: people. According to an analysis published this week, DeepMind's hiring ratio has collapsed from 12 new researchers for every departure in mid-2023 to roughly 2-to-1 by late 2026, and alumni have been launching rival startups within walking distance of the lab's London offices.

That's not a lab losing talent at random — it's the predictable friction of a mission change. Researchers who joined for open-ended inquiry leave when the job becomes shipping quarterly releases. For developers and founders building on Gemini, the new tempo is probably good news in the short term. Whether DeepMind can keep its research edge while running itself like a product factory is the open question Kavukcuoglu will be judged on.

What to watch#

• The first public build. Watch for who gets the early post-training version — limited preview, partner testing, or general availability — and how quickly iterations follow.
• Agentic coding benchmarks. If Gemini 4's headline demos are about autonomous coding rather than chat quality, that's Google conceding where the race is being run.
• The cadence promise. A monthly Gemini rhythm would be a genuine organizational feat for DeepMind; the first slip will tell you whether it's real.
• The talent line. If the 2-to-1 hiring ratio keeps falling while rivals fund ex-DeepMind startups, the "Gemini factory" may find itself short of the people who built the first three generations.

Google spent most of 2026 quietly rebuilding its release machinery. Gemini 4 will show whether the machine works.