Gemini 4 Enters Post-Training: What Google's Timeline Signals

Google DeepMind's next flagship model has moved into its final training stage before release, and its new operational chief is signaling he wants to ship it sooner rather than later. Speaking at The Information's AI Agenda Live Summit on September 23-24, 2026, Koray Kavukcuoglu, elevated to Senior Vice President of Google DeepMind on August 12, 2026, confirmed that Gemini 4 has entered post-training and said Google intends "to roll out an early post-training version as soon as possible, because we've already seen promising results and are very excited." He gave no firm date, only that the target is "much earlier" than the end of 2026.

Google

Gemini

OpenAI

Anthropic

Google, OpenAI and Anthropic are now racing on overlapping release calendars, and Kavukcuoglu's comments are a signal about where Gemini 4 fits into that calendar.

That is a statement of intent, not a shipped product, and the distinction matters more than usual here. Google's Gemini line has not had a flagship release since Gemini 3 Pro in November 2025, and a planned Gemini 3.5 Pro was teased and then quietly abandoned after missing several internal deadlines, with Google shifting its 2026 releases to a run of Flash-tier updates instead. Kavukcuoglu's remarks are the clearest public signal yet that a genuine flagship successor is close enough to talk about, and the way he talked about it, an early, rougher version shipped fast rather than a single polished launch, is itself a choice worth unpacking.

Aug 12, 2026Kavukcuoglu named SVPTook operational lead of Google DeepMind

Sept 23-24, 2026AI Agenda Live SummitWhere the post-training update was confirmed

34.0Gemini 3.6 Flash scoreIntelligence Index v4.3.2, per The Information

57.6Claude Opus 5.5 scoreSame benchmark, current leader as reported

~41%Gap to the leaderComputed from the two scores above

What "post-training" actually means

Building a frontier language model happens in two broad phases, and the terminology is doing real work here. Pre-training is where a model first learns language, facts and reasoning patterns, by predicting text across an enormous corpus of data. It is the most compute-intensive phase and can run for months on tens of thousands of accelerator chips. What comes out the other end is a raw base model: capable in a statistical sense, but not yet a usable assistant. It has no consistent sense of how to follow an instruction, decline an unsafe request, or format an answer the way a person actually wants it.

Post-training is what turns that raw model into a product. It typically layers supervised fine-tuning on curated examples, reinforcement learning from human and AI feedback to sharpen judgment and instruction-following, and safety testing to catch harmful or unreliable behavior before it reaches users. As one report on Kavukcuoglu's comments put it, post-training is where a lab refines "the base model through human feedback, reinforcement learning, and specialized data to enhance reasoning, safety, and reliability." It is also the stage where a model's actual personality and reliability, the parts users experience directly, get set. A model can pre-train well and still ship badly if post-training is rushed or under-resourced, which is exactly why an executive publicly discussing this specific stage, rather than staying silent until a launch is ready, is notable.

Where Gemini 4 sits in the release pipeline

The general shape of a frontier model release, with Gemini 4’s reported stage highlighted. Illustrative, not Google’s own internal roadmap.

Pre-training

The base model learns general language and reasoning patterns from a large corpus of text, code and other data. Compute-heavy, can run for months.

Post-training

Kavukcuoglu confirmed Gemini 4 is here: supervised fine-tuning, reinforcement learning from human and AI feedback, and safety testing shape the base model into an assistant.

Early rollout

The stage Google says it wants to reach "as soon as possible": an early post-training version shipped ahead of a fully polished release, then iterated on with real usage.

General release

A stable, broadly available model. Not yet dated for Gemini 4; Kavukcuoglu said only that it is targeted "much earlier" than the end of 2026.

Source: Koray Kavukcuoglu, SVP of Google DeepMind, at The Information’s AI Agenda Live Summit, September 23-24, 2026.

Why say this in public, and why now

Announcing a mid-pipeline milestone rather than a launch date is an unusual amount of transparency for a frontier lab, and it reads as a deliberate signal aimed at three audiences at once: investors watching Alphabet's AI narrative, developers deciding which model family to build against, and rivals watching the calendar. Google's own reporting places Gemini 3.6 Flash, its current frontier-tier model, at 34.0 on the Intelligence Index v4.3.2 benchmark, trailing OpenAI's GPT-6 Sol at 47.5, GPT-6 Astra at 52.7, and Anthropic's Claude Opus 5.5 at 57.6, a gap of roughly 41% behind the reported leader. Against that backdrop, confirming that a genuine successor is in its final stage, and specifically saying Google wants to skip the wait for a fully polished release, functions as reassurance that the gap is closing rather than a static state of affairs.

Announcing a mid-pipeline milestone rather than a launch date is an unusual amount of transparency for a frontier lab.
On why Kavukcuoglu's comments matter

The competitive backdrop explains the urgency. OpenAI has been shipping its GPT-6 family since early September, with GPT-6 Astra launching September 3, 2026, and Anthropic released Claude Fable 5.1 on September 1, 2026, on top of the Opus 5.5 release referenced above. A pricing war among frontier models broke out in the same window, with Grok 4.7, Opus 5.5, GPT-6 Sol and other models all repricing on September 21-22, 2026. Every major lab is shipping on an overlapping, compressed schedule, and staying quiet about Gemini 4's status for another few months would have let that narrative run unchallenged.

It is also worth remembering that this compression cuts both ways. When Gemini 3 launched in November 2025, its benchmark performance reportedly surprised OpenAI badly enough that CEO Sam Altman declared an internal "code red" in a December 1, 2025 memo, redirecting resources toward ChatGPT's quality and reliability. Google has been on the trailing side of this race for most of 2026, but the race itself has already flipped direction once in the space of a year. That history is the reason "we've already seen promising results" is a meaningful thing for Kavukcuoglu to say in public rather than empty confidence: Google has, at least once recently, actually had the model that made a rival panic.

A Google campus sign at the Googleplex in Mountain View, California — A Google campus sign in Mountain View, California, photographed in 2022. Gemini 4 is developed inside Google DeepMind, the research organization Kavukcuoglu now runs operationally within Google; this photo shows Google's corporate campus, not a DeepMind facility or anything specific to Gemini 4.

The risk in pre-announcing, and what to watch

Talking about a model before it ships carries real downside, and Google's own recent history illustrates it. Gemini 3.5 Pro was teased publicly and then missed enough internal deadlines that Google effectively abandoned it without ever formally canceling it, routing its 2026 roadmap through Flash-tier updates instead. If Gemini 4's "much earlier than year-end" target slips the way 3.5 Pro's did, the story flips from "Google is closing the gap" to "Google keeps promising a flagship it cannot ship," which would compound rather than relieve the pressure Kavukcuoglu is currently trying to manage. Pre-announcing a milestone this specific is a bet that the model is far enough along to be worth the exposure, made by an executive six weeks into the job whose credibility is now partly staked on the outcome.

The other detail worth watching is what "early post-training version" actually means when it ships. Kavukcuoglu has said internal testing is already using Antigravity, Google's agentic development environment, and framed Gemini 4's focus around coding, autonomous agents and long-horizon agentic workflows rather than raw chat quality. A model released straight out of early post-training, before the later rounds of safety tuning and reliability polish a fuller cycle would normally include, is likely to be capable but rough around the edges. Whether Google frames that honestly as a preview, or lets it be read as a finished flagship, will say a lot about how much the release timeline is being driven by genuine readiness versus competitive pressure to have something to point to.

None of this changes the practical reality for teams building on these models today: the frontier is moving, the ranking between Google, OpenAI and Anthropic has flipped before and can flip again, and no single lab's roadmap is a safe thing to build a product around exclusively. That is the whole case for staying model-agnostic. Metir AI routes a task to whichever frontier model actually performs best for it, so when Gemini 4 does ship, in whatever form it takes, it becomes another option to route to rather than a migration a team has to plan around.

Sources:

Image credits

Hero and in-body figure: a Google campus sign at the Googleplex in Mountain View, California, photographed in 2022, via Wikimedia Commons, licensed under CC BY-SA 4.0. Reviewed before publication; shows Google's Mountain View corporate campus, not a Google DeepMind-specific building or anything depicting Gemini 4 itself, as the caption states.