Post-Training and Counting: Google Confirms Gemini 4 Is Weeks, Not Months, Away

For most of 2026, the question hanging over Google has been uncomfortably simple: where is the next flagship Gemini? This week we finally got an answer with a date attached — or at least a season. In his first media appearance since taking over as chief of Google DeepMind, Koray Kavukcuoglu confirmed that Gemini 4 has entered post-training, the final refinement phase before a public release, and that Google intends to ship it “as soon as possible” — and, he hopes, “much earlier” than the end of 2026.

The comments, made to The Information at its AI Agenda Live Summit and corroborated by The Verge and 9to5Google, are the strongest signal yet that Google’s long-awaited frontier model is weeks rather than months away. For a company that has spent nearly a year watching OpenAI and Anthropic trade flagship releases, it amounts to a public commitment to re-enter the race on a clock.

What “post-training” actually means

Post-training is the stage where a raw base model — the artifact that emerges from the enormous pre-training run that began on July 21, 2026 — is turned into a product. Behavior is refined through reinforcement learning and fine-tuning, safety testing and guardrails are implemented, and internal dogfooding ramps up. Kavukcuoglu confirmed that Google is already testing Gemini 4 internally, using it to power Antigravity, the company’s agentic coding environment.

That detail matters more than it might first appear. Testing a frontier model inside a real agent product — rather than a chat window — signals that Google is optimizing Gemini 4 for agentic workloads from the start, not retrofitting agent capabilities after the fact. When Kavukcuoglu was asked about AGI, he reframed the conversation entirely: “the conversation is more about are we able to build intelligent agents that we can trust.” That is a telling shift for a lab that spent 2025 talking about benchmark ceilings.

His full quote on the timeline: “Our intention is to, like — as soon as possible — to release an early post-training output because we see the results and we are excited… [to] continue the fast-paced iterations.”

Two things stand out. First, the phrase “early post-training output” suggests Google may stage the release — shipping an initial checkpoint quickly and iterating in public, the same cadence the company has honed with its Flash line. Second, “we see the results and we are excited” is the kind of unprompted enthusiasm that labs usually suppress; internal evaluations of Gemini 4 are evidently strong enough that leadership is willing to bind the company’s credibility to a fast ship date.

A year of watching from the sidelines

The context for this urgency is brutal. Google’s last true flagship release was Gemini 3 Pro in November 2025, followed by Gemini 3.1 Pro in February 2026. Since then, the frontier has moved without it: OpenAI has shipped GPT-6 and this week’s GPT-6 Sol and Luna variants at aggressive price points, while Anthropic released Claude Opus 5.5 — dubbed “Fable power at Opus prices” — plus its Mythos series. Each of those models outperforms Gemini 3 by measurable margins.

Google, meanwhile, spent the year on a conspicuous detour. A Gemini 3.5 Pro update was teased over the summer and promised by CEO Sundar Pichai at I/O in May for a June arrival — and never shipped. Kavukcuoglu admitted the company “took a little bit of a step back” to focus on faster, cheaper Flash models instead, most recently Gemini 3.8 Flash. The Flash line has been genuinely successful at the low end, but it did nothing to answer the perception that Google had no answer at the frontier.

The leadership shakeup amplified the scrutiny. Demis Hassabis, DeepMind’s co-founder and long-time CEO, stepped back from daily operations in August — notably, in the same note announcing that “great progress” was being made on Gemini 4. Kavukcuoglu inherited both the model and the expectation gap. When pressed on whether Google had fallen behind, his answer was defiant: “I have the utmost trust in the team… In my mind, it’s a certainty that we are always gonna be at the frontier.”

Why the timing is plausible

Skeptics will note that Google has cried “coming soon” before — the phantom Gemini 3.5 Pro being the freshest example. But the post-training signal is structurally different from a teaser. Pre-training timelines are long and uncertain; post-training is a bounded engineering phase with a known end state. The Information’s reporting that Google is “nearing release” aligns with the July 21 pre-training start date: roughly a nine-week pre-training run followed by post-training beginning in September is consistent with an October or November ship.

There is also competitive logic. Every week Gemini 4 slips, OpenAI’s DevDay (with GPT-6 Cyber walking out of preview) and Anthropic’s momentum claim more developer mindshare. Google’s distribution advantages — Android, Workspace, AI Studio, and the Gemini app’s billion-user base — are only an advantage if the underlying model is competitive.

What to watch

No specific launch date, market, or channel has been announced, and Google has published no model page, API identifier, or pricing. Watch for three tells that a release is imminent: benchmark numbers leaking from internal evaluations (rumors already place Gemini 4 ahead of GPT-5.6 Sol and Claude Fable 5, though these are unverified), an Antigravity update shipping with Gemini 4 inside, and a sudden cadence of safety documentation — post-training is when guardrail testing papers appear.

Whenever it lands, Gemini 4 will be the first flagship model of the post-Hassabis era at DeepMind, tested in an agent product before it ever reached a chat box. If the “early post-training output” ships as promised, 2026’s narrative of Google playing catch-up may invert quickly — and if it doesn’t, Kavukcuoglu’s “certainty” quote will follow him for a long time.