From Pre-Training to Post-Training in Nine Weeks: Google's New DeepMind Chief Vows Gemini 4 Will Ship in 2026

Nine weeks. That is roughly how long it took Google to take Gemini 4 from the start of what it called its “most ambitious pre-training run yet” to the refinement stage that precedes a launch. On September 24, Google DeepMind’s new chief, Koray Kavukcuoglu, confirmed that Gemini 4 has formally entered early post-training — and, in a striking pledge for a company that usually refuses to talk about ship dates, he said the next frontier model should arrive “much earlier” than the end of 2026.

What happened

The announcement came through a round of interviews and coverage published on September 24, led by The Verge’s report that Gemini 4 is “almost ready” according to Kavukcuoglu, who took over the top job at Google DeepMind in the sweeping August reorganization that saw Demis Hassabis step aside into a new role. The Verge’s takeaway was blunt: the new frontier model should arrive long before the end of 2026.

Kavukcuoglu described the model as being in an early phase of post-training — the stage where a raw base model is refined with techniques like reinforcement learning and fine-tuning until it behaves reliably enough to face customers. And he signaled an aggressive rollout philosophy, telling reporters that the intention is to share an early post-training version “as soon as possible,” because the team believes rapid iteration in public beats a long, quiet polish.

That is a meaningful departure from Google’s usual cadence. When the company confirmed the Gemini 4 pre-training run on July 21 — tucked into the launch post for Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber — it confirmed nothing else: no model sizes, no benchmarks, no pricing, no context window, no API details, no preview program. Nearly a month into post-training, that is still true. What Google has offered, for the first time, is a timeline commitment.

Why nine weeks is the real headline

The compressed schedule matters more than the vague promise. Gemini 4’s pre-training run began on July 21, 2026. By late September, the model was already in post-training. As Yahoo Finance’s analysis noted, a transition from pre-training to post-training in approximately two months is a dramatically compressed timeline by the standards of frontier-scale training runs, which have historically stretched across many months of compute, evaluation, and re-runs.

There are two ways to read that speed. The generous reading: Google has industrialized its training pipeline to the point where a next-generation frontier model can move from raw compute to refinement in a single quarter, and Gemini 4’s “most ambitious” run was efficient enough to hand off early. The skeptical reading, articulated by analysts as recently as early August, was that a large new model shipping in 2026 “would have to be bizarrely fast” — and that the realistic outcome was a smaller or mid-generation release dressed up with a big number.

Kavukcuoglu’s comments are effectively a bet on the generous reading. He is staking his first months as chief on shipping a full frontier-generation model within roughly five months of starting its pre-training run.

A new chief with a different answer to the AGI question

The subtext of the announcement is a change in what Google DeepMind optimizes for. The Decoder’s coverage made the cultural shift explicit: DeepMind was built to chase AGI, but its new chief mostly just wants Gemini 4 out the door. Where Hassabis spent years framing the lab’s mission around artificial general intelligence — a 5-to-10-year horizon, research breakthroughs still to come — Kavukcuoglu has repeatedly deflected the AGI question in favor of shipping, iteration, and developer traction.

That priority shows in where he has spent his public appearances. In a wide-ranging podcast conversation with Google’s Logan Kilpatrick shortly before the announcement, Kavukcuoglu discussed frontier models, coding agents, and why “the frontier is all that matters” — a framing that treats benchmark leadership and product velocity as the same problem.

The competitive clock is loud

The urgency is not mysterious. In the span of a single late-September week, OpenAI shipped GPT-6 Sol and Luna, and Anthropic released Claude Opus 5.5 — a model positioned to match its Fable-class flagship at a sharply lower API price. Google’s current top-line models, the Gemini 3.x family, are increasingly bracketed by rivals whose newest releases arrive on a roughly 18-day cadence across the major labs, according to one widely shared tally of 2026 release data.

The sorest spot is coding. As The Information reported in April, DeepMind researchers themselves viewed Anthropic’s coding tools as exceeding Gemini’s code-writing abilities, prompting Google to assemble an internal strike team focused specifically on coding model performance. Coding agents are the killer app of this cycle — the place where frontier capability converts directly into developer revenue and enterprise contracts. If Gemini 4 does not close that gap, the rest of its benchmark sheet will matter much less.

What we still don’t know

It is worth repeating what Google has not said. There is no confirmed release date — only “well before year-end” and “much earlier.” There are no announced specs: parameter counts, context length, pricing tiers, multimodal scope, and API availability all remain blank. Safety evaluation and a staged rollout still stand between “early post-training” and general availability, and coverage of the announcement noted that safety work continues with no fixed date attached.

The staged-rollout caveat is not boilerplate. Frontier models that ship on compressed timelines tend to land first as experimental or trusted-tester previews, with general availability following weeks later. An “early post-training version” shared with developers before the end of October, followed by a broader launch in November or December, is the most plausible shape of what Kavukcuoglu is promising.

Why it matters

Google is the only one of the three frontier labs that also owns the distribution: Search, Android, Workspace, and Chrome. When Gemini 4 ships, it will not arrive as a standalone chatbot competing for attention — it will propagate across products with billions of users, which is precisely why the company can afford to talk about timelines this aggressively. A strong Gemini 4 arriving early resets the competitive calendar for OpenAI and Anthropic; a weak or delayed one hands them a quarter of uncontested momentum.

Nine weeks from pre-training to post-training is either a sign that frontier training has become an industrial process — or a sign that Google is about to ship a model it did not have time to finish. By the end of December, we will know which.