Google DeepMind Confirms Gemini 4 Has Entered Post-Training, Aiming to Launch Well Before Year-End

On September 24, 2026, Google DeepMind's newly appointed SVP Koray Kavukcuoglu confirmed at The Information AI Agenda Live Summit that Gemini 4 has completed pre-training and entered the early post-training stage.

Fact Recap

According to information from the summit, Gemini 4 is currently in the post-training stage, which includes safety testing and the implementation of guardrails. Google has already used Gemini 4 to power its internal product Antigravity for testing. Kavukcuoglu said the team intends to release an early post-training version as soon as possible, because they felt excited after seeing the results, and wants the model's release to come well before the end of 2026. Google's previous major flagship model release was Gemini 3 Pro in November 2025, followed by Gemini 3.1 Pro in February 2026, with series updates such as Flash in between, most recently Gemini 3.8 Flash.

How It Works

The post-training stage is a key step in moving a model from pre-training to actual deployment, focusing on behavior refinement, safety validation, and guardrail setup. Google's choice to first use Gemini 4 to power the Antigravity product internally indicates that this stage has already reached a testable state. Kavukcuoglu mentioned that rapid iteration will continue, which contrasts with the recent backdrop of Anthropic and OpenAI launching flagship model updates. Google previously briefly considered Gemini 3.5 Pro but ultimately did not release it, shifting its focus to the Flash series instead, and now the emphasis is turning to Gemini 4.

Industry Impact

Against the backdrop of rapid iteration by competitors, the advancement of Gemini 4 could affect Google's positioning at the AI frontier. Kavukcuoglu said he has full trust in the team and believes Google will always be at the frontier. An earlier release date could change market expectations for the pace of Google's AI products, while positive internal test results would provide a basis for subsequent product integration.

Strategic Assessment

[This section is analysis rather than fact] Based on available information, Google setting Gemini 4's release at a point well before the end of the year may be intended to respond to competitive pressure and maintain its pace of iteration; if the early post-training version is released smoothly, it could help narrow the gap with the leaders, but the specific effect still needs to be verified by subsequent actual performance.