Post-Training Has Begun: Google's Gemini 4 Enters the Final Stretch Early

The race to ship the next frontier model just accelerated on schedule — Google’s. On Wednesday, September 23, the new top leader of Google DeepMind, Koray Kavukcuoglu, told The Information that Gemini 4 is in the early stages of post-training, the development phase in which a base AI model is refined to behave reliably. An initial version is expected “as soon as possible,” with the full model anticipated well before the end of 2026.

That is a compressed timeline by any modern standard. Google first confirmed Gemini 4 pre-training on July 21, 2026, in a passing mention buried at the bottom of a Gemini 3.6 Flash launch post — an unusual place for the biggest bet in the company’s history. Two months later, the model has already cleared the most compute-intensive phase of its development and moved into refinement. For comparison, the gap between Gemini generations historically ran longer; the industry’s new normal, demonstrated by OpenAI and Anthropic this month, is overlapping training pipelines and a release roughly every quarter.

Why This Matters

Gemini 4 arrives at a pointed moment for Google. In August, Reuters reported that the company had delayed the flagship by two months after internal testing showed its performance continued to lag behind internal targets. The same month, Google executed its largest AI leadership reshuffle in years: Demis Hassabis stepped down as CEO of Google DeepMind to become chairman and chief scientist of Alphabet, handing operational leadership to Kavukcuoglu — a veteran of the Gemini model program and one of the most senior technical figures in the company.

The Gemini 4 timeline is, in effect, the first big public test of the new org chart. Kavukcuoglu’s comments to The Information are the clearest signal yet that the reorganization was about speed: cutting decision latency between research and shipping, and closing the gap between DeepMind’s internal ambitions and what developers actually see in the API.

The Context of the Competitive Race

The timing puts Google in direct response mode against two rivals who have already moved:

• OpenAI shipped GPT-6 Sol and GPT-6 Luna on September 22, cutting token prices aggressively (Sol at $2/$10 per million input/output tokens, Luna at $0.10/$0.50) and rolling them into ChatGPT Work and Codex for most paid accounts within hours.
• Anthropic released Claude Opus 5.5 the same day, matching Fable-class performance at 40% lower cost than Opus 5 ($4/$20 per million tokens), with improved safeguards.

Both releases pressed hard on the two axes that matter most to enterprise buyers right now: agentic reliability and price. Google’s counter has so far been the Gemini 3.x line — including Gemini 3.8 Flash TTS, 3.8 Live, and Extended Thinking variants shipped through September — which expanded surface area without moving the frontier flagship. Gemini 4 is the first model designed from the start under the new leadership to answer the price-performance question at the flagship tier.

What We Actually Know — and Don’t

Confirmed facts about Gemini 4 remain sparse, which is itself part of the story:

• Pre-training began by July 21, 2026 (Google’s own confirmation, via the Gemini 3.6 Flash post).
• It is now in early post-training (Kavukcuoglu, The Information, September 23).
• An initial version is planned “as soon as possible,” with full release expected significantly earlier than the end of 2026 (multiple outlets following The Information’s report).
• Sundar Pichai has described it as significantly larger than anything Google has trained before — though no parameter count, context window, pricing, or benchmark numbers have been published.
• The stock market has already voted: shares of Alphabet moved on the report within hours.

What we don’t know: benchmark performance, API pricing, whether there will be a staggered rollout (initial version vs. full model), and how the open-weight Gemma line will inherit from it. Analysts covering the story have been careful to separate confirmed facts from speculation, and Google has published no release date, model ID, or API documentation.

Analysis: The Overlap-Cadence Era

The most significant signal in this story isn’t any single spec — it’s the structure of the timeline. Google is now training multiple generations in parallel and targeting a faster release cadence, a strategy it has leaned into publicly since mid-2026. Pre-training for Gemini 4 started while Gemini 3.8 variants were still shipping; post-training begins while OpenAI’s and Anthropic’s September flagships are days old.

This overlap-cadence model changes what “release” means. When generations overlap, an initial Gemini 4 version landing early doesn’t need to beat every rival on every benchmark — it needs to reset the frontier reference point and lock developers into the ecosystem before the next competitor moves. The Information’s report that the team plans to launch an initial version “as soon as possible” suggests Google is optimizing for exactly that: presence at the frontier, continuously, rather than a single dramatic unveiling.

There are risks. The August delay reportedly came from performance lagging internal targets — rushing post-training to hit an early window is precisely when alignment and reliability work gets squeezed. Kavukcuoglu’s framing (“early days” of post-training) leaves room for the timeline to slip again, and Google’s rivals have demonstrated they will punish any hesitation with pricing.

But the direction is clear. Under its new leadership, Google DeepMind is no longer managing to a calendar; it’s managing to a race. And by the company’s own account, Gemini 4 is closer than anyone outside the building expected.