Google moves Gemini 4 into post-training, testing it inside Antigravity

Why it matters

Gemini 4 is moving from training into behavioral tuning and internal agent testing. Google's next challenge is to show developers what it can do outside its own systems, and when they can try it.

Google has moved Gemini 4 into post-training and is testing it internally in Antigravity, its agentic software-development platform, according to 9to5Google's report on remarks by Google DeepMind chief Koray Kavukcuoglu at The Information's AI Agenda Live Summit on September 23rd. Kavukcuoglu said Google wants to release an early post-training version "as soon as possible" and hopes to make Gemini 4 available well before the end of 2026. He gave no launch date.

The status update puts a practical test ahead of the usual frontier-model scorekeeping: whether Google's next flagship can power agents people will trust to carry out work. In an accessible reproduction of The Information interview, corroborated by The Decoder's report, Kavukcuoglu said, "The conversation is more about are we able to build intelligent agents that we can trust." That framing fits Google's decision to test Gemini 4 in Antigravity, where the model is being used in a software-development product rather than discussed only as a research milestone.

The leader behind the release

Kavukcuoglu is a research leader, not Gemini 4's startup founder. Google says he previously served as Google DeepMind's vice president of research and founded its deep-learning team. The group helped pioneer DQN, a reinforcement-learning system, and WaveNet, a neural audio-generation model. His current role puts him at the junction of model research and Google's product deployment, making his comments a window into what Google wants this model to do.

His description of post-training is more revealing than the phrase "nearing release" on its own. Google is refining the model's behavior, conducting safety tests and implementing guardrails before wider availability. An early post-training output would let Google continue iterating with a version that has already moved beyond initial training, while the work of tuning and testing continues. That is a stated intention, not a schedule or a promise of general availability.

Kavukcuoglu said Google had "taken a little bit of a step back" from Gemini 3.5 Pro to focus on Flash models, and that its current focus is Gemini 4, according to 9to5Google's account of the interview. Google has not said whether Gemini 3.5 Pro will still ship. Google's sequence of model names and releases has made the flagship timeline harder to read: Gemini 3 Pro arrived in November 2025, Gemini 3.1 Pro followed in February 2026, and Google has since issued multiple Flash updates.

An agent test, not a benchmark reveal

Testing Gemini 4 in Antigravity gives Google an internal use case while it works on the model. Google has positioned Antigravity as a place to build with AI agents; using a not-yet-released flagship there lets its teams assess how the model behaves in a software workflow. It does not establish that Gemini 4 will outperform competitors at coding, or that external developers will get access through Antigravity first. Google has disclosed no Gemini 4 benchmark results or release format.

Google has put real scale behind its developer and agent products. On its July 22nd earnings call, CEO Sundar Pichai said more than 9 million developers were building monthly with Google's models across APIs and key developer products, and that model APIs were processing about 22 billion tokens per minute. Pichai also said Antigravity had more than 2.4 million weekly active users. Those are company-reported figures, and they describe Google's existing business, not Gemini 4 adoption. They help explain why model quality in an agentic coding environment matters to Google: it already has developer products and distribution channels in which a stronger model could be used.

Google has issued multiple Flash updates since Gemini 3.1 Pro, most recently Gemini 3.8 Flash. The Gemini 4 update shifts the emphasis from Flash iteration toward the next flagship, while leaving the same operator question open: what can developers actually test, and on what terms?

Kavukcuoglu said Google remains confident it will stay at the frontier, despite questions about whether it has fallen behind rivals such as OpenAI and Anthropic. The claim will be judged when Gemini 4 can be compared and used outside Google's internal systems. For now, the concrete progress is a model in post-training, an internal Antigravity test, and leadership's stated intention to release an early version quickly. Google has disclosed no benchmark results, developer access terms or release date.