Gemini 4: Google DeepMind's New Chief Wants It Out 'As Soon As Possible'

Google DeepMind's new chief, Koray Kavukcuoglu, says the Gemini 4 ASAP release is coming — the model is in post-training and Google wants an early build out "as soon as possible," much earlier than year-end, in his first public appearance since taking over the lab on August 12, 2026.

The Quote That Changed the Timeline

On September 23, at The Information's AI Agenda Live Summit, Google DeepMind's newly appointed chief, Koray Kavukcuoglu, said something that narrowed the window on the Gemini 4 ASAP release.

"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," Kavukcuoglu told The Information.

Three words — "as soon as possible" — and a reason — "we see the results and we are excited."

That is a different kind of signal than the one Google sent when Gemini 4 entered post-training. Post-training entry is a milestone. "As soon as possible" is a timeline posture. Together, they describe a model that is not just moving toward release but moving toward release with urgency.

Kavukcuoglu reinforced the point in a separate interview with 9to5Google on September 24: Google wants to ship an early build "much earlier than year-end." No calendar date. No month. But the direction of travel is clear — earlier than the end of 2026, and sooner than the outside world had been assuming.

Who Kavukcuoglu Is and Why His First Public Appearance Matters

Kavukcuoglu was elevated to Senior Vice President of Google DeepMind on August 12, 2026. The AI Agenda Live Summit, on September 23-24, was his first public appearance in the role.

That matters for two reasons.

First, it means the "as soon as possible" framing for the Gemini 4 ASAP release is coming from the person now responsible for the lab's flagship model roadmap — not from a spokesperson, not from a leaked slide, not from an anonymous source. It is the new chief, in his first public setting, setting a tone.

Second, it means we are seeing Kavukcuoglu establish his public position on the questions that have defined Google DeepMind's public identity. His answer to the AGI question, in this first appearance, was notably understated: he called the AGI framing "not the right conversation" and said trustworthy agents matter more. That is a posture shift from the Hassabis era, where AGI was a live and explicit topic of public discussion.

Kavukcuoglu is signaling, in his first public appearance, that the next chapter of Google DeepMind is less about declaring AGI and more about shipping capable, trustworthy models — and that the Gemini 4 ASAP release is the first concrete expression of that posture.

What Post-Training Means in This Context

Post-training is the phase in which a base model's behavior is refined ahead of public release. It is where human feedback, safety tuning, capability shaping, and deployment preparation happen.

A model in post-training is not a model in training. The base capability is largely set. What remains is the work of making the model behave the way the lab wants it to behave when it reaches users — including, in Google's case, the work of integrating it into products.

The fact that Gemini 4 is already running internally in the coding tool Antigravity tells you something about where Google is in the process. A model that is being used internally by the people building it is a model that is past the early stages. It is being tested, used, and presumably evaluated against the internal benchmarks that determine whether it is ready to ship.

The Three-Way Race, Updated

The frontier model race has a shape right now, and the Gemini 4 ASAP release changes one corner of it.

OpenAI has GPT-6 Astra live — rated "Critical" on ExploitBench — and is previewing GPT-6 Cyber at DevDay on September 29. GPT-6 Sol and Luna are shipping. OpenAI's models are in users' hands and in production.

Anthropic has Claude Opus 5.5 and Claude Fable 5.1 shipping, with Fable 5.1 being the cost leader in its class and Opus 5.5 landing in GitHub Copilot. Anthropic's models are in users' hands and in production.

Google has Gemini 3.8 and Gemini 3.8 Flash TTS in the market, and Gemini 4 in post-training — with the new chief saying Google wants an early build out "as soon as possible" and "much earlier than year-end."

The gap between Google and the other two labs is real and measurable. OpenAI and Anthropic are shipping. Google is preparing to ship. The question that Kavukcuoglu's quote answers is not whether Google will close that gap — it is how fast.

A "as soon as possible" release of an early post-training build would be a concrete step toward closing that gap. It would not close it entirely. But it would put a Google flagship in users' hands before the end of 2026, which is earlier than the year-end expectation that had been the working assumption.

What "As Soon As Possible" Does and Does Not Mean

"As soon as possible" is not a date. It is a direction.

It does not tell you whether the Gemini 4 ASAP release happens in October, in November, or in December. It does not tell you whether the early build is a full flagship release or a partial one. It does not tell you what the benchmarks will look like or how Gemini 4 will compare to GPT-6 Astra and Claude Opus 5.5 on the metrics that matter.

What it does tell you is that Google is not inclined to hold the model back for a carefully orchestrated year-end launch event. The results are good enough, in the lab's assessment, to warrant an early release — and the lab is excited about those results.

That is a meaningful data point for anyone tracking the model race. It says Google sees something in Gemini 4 worth showing sooner rather than later.

The Internal-Use Signal

The detail that Gemini 4 is running internally in Antigravity is underappreciated in most of the coverage, but it is one of the most concrete pieces of evidence we have about the model's readiness.

A model that is integrated into an internal coding tool is a model that is being used for real work by the people building it. That is not the same as a model that is ready for public release — internal use is a lower bar than public launch. But it is a higher bar than a model that exists only in evaluation harnesses.

It suggests Gemini 4 is at a stage where Google's own engineers are relying on it. That is a meaningful step on the path to a public release, and it is consistent with an "as soon as possible" timeline for the Gemini 4 ASAP release.

The AGI Posture Shift

Kavukcuoglu's dismissal of the AGI conversation as "not the right conversation" is worth sitting with, because it is a real departure from how Google DeepMind has talked publicly under Demis Hassabis.

Hassabis has been explicit about AGI as a goal and a topic of discussion. Kavukcuoglu, in his first public appearance, is saying the AGI framing is not the right one — and that the right conversation is about trustworthy agents.

That could be a substantive shift in how Google DeepMind thinks about its mission. It could also be a messaging shift — a way of talking about the same work in a different register. We do not yet know which.

What we do know is that the public voice of Google DeepMind has changed, and the change is visible in the Gemini 4 ASAP release story. The new chief is less interested in declaring AGI and more interested in shipping a model that works.

The Context Google Is Operating In

The "as soon as possible" quote lands in a particular moment for Google.

Gemini 3.8 Flash TTS launched on September 26 — two voice models built for scripting and volume. That is a product release in the Gemini family, and it signals that Google is actively shipping Gemini-branded products even as Gemini 4 moves through post-training.

The White House has asked OpenAI and Anthropic to delay sharing models with UK testers until U.S. review — a governance dynamic that affects the whole field, including Google. The EU has invited frontier labs to "pace the frontier," a posture that cuts both ways. And Google DeepMind, under new leadership, is making its first public moves.

In that context, an early Gemini 4 ASAP release is not just a product decision. It is a statement about where Google thinks it stands in the race and how aggressively it wants to compete.

What We Do Not Know

There are real limits to what the current reporting tells us.

We do not have a release date, or even a month. "As soon as possible" and "much earlier than year-end" are directions, not commitments.

We do not have Gemini 4's benchmark scores. We know the model is in post-training and running internally. We do not know how it performs relative to GPT-6 Astra, Claude Opus 5.5, or Claude Fable 5.1 on the public benchmarks that matter.

We do not know whether the early build will be a full release or a partial one, or what Google's definition of "early post-training output" actually means in practice.

We do not know how the new leadership posture — less AGI talk, more trustworthy agents — will translate into the model's design, its safety profile, or its deployment strategy.

The Watch Points

Four things to watch as the Gemini 4 ASAP release story develops:

First, a concrete date or month. "As soon as possible" will become a lot more interesting when it acquires a calendar anchor. If Google commits to a release window, that is a signal that the post-training work is on a known track.

Second, benchmark leaks or disclosures. A model in post-training will eventually have scores, and those scores will tell you where Gemini 4 stands relative to GPT-6 Astra and Claude Opus 5.5.

Third, an official Google announcement. The current reporting is based on Kavukcuoglu's comments to The Information. An official release from Google would be a different kind of signal — one that says the company is ready to stand behind the model publicly.

Fourth, how the AGI posture shift plays out in the product itself. If Gemini 4 is, in practice, less about AGI claims and more about agentic capability and trustworthiness, that would be a concrete expression of Kavukcuoglu's stated orientation — and a meaningful difference from the way Google DeepMind has talked under Hassabis.

The Relationship to Post 50

AIPress covered Gemini 4's entry into post-training on September 27, in Post 50. That post established the milestone and framed it as a concrete signal in the three-way race.

The "as soon as possible" quote is new data that Post 50 did not include. It strengthens the post-training story by adding a timeline posture to the milestone. It also introduces Kavukcuoglu's first-public-appearance framing — the AGI posture shift, the Antigravity internal use, the "we see the results and we are excited" rationale — that Post 50 could not have captured because the summit had not happened yet.

This post is best read as a companion to Post 50: same model, same phase, new details, and a sharper read on what Google's timeline actually looks like for the Gemini 4 ASAP release.

The Bottom Line

Google DeepMind chief Koray Kavukcuoglu says Gemini 4 is in post-training and that Google intends to release an early build "as soon as possible" — much earlier than the end of 2026 — because the lab "sees the results and" is "excited."

It is Kavukcuoglu's first public appearance since taking over Google DeepMind on August 12, and he used it to set a faster timeline than the year-end assumption and a different public posture than his predecessor — less AGI talk, more trustworthy agents.

Gemini 4 is already running internally in the coding tool Antigravity. Gemini 3.8 Flash TTS shipped on September 26. And Google, under new leadership, is signaling that it wants to close the gap to OpenAI and Anthropic sooner rather than later.

We do not have a date. We do not have benchmarks. We do not have an official announcement.

But we have an urgency signal from the person in charge — and in the frontier model race, that is something.