Gemini 4 has reached early post-training, Google DeepMind leader Koray Kavukcuoglu said at The Information’s AI Agenda Live Summit on September 23, 2026. He hopes to release an initial version “much earlier” than the end of the year and improve it through fast iterations. That is a development update and a release ambition—not a launch announcement.
19
13
What early post-training means for Gemini 4
Post-training is the stage in which a base model’s behavior is refined ahead of release, including work on reliability, safety testing and guardrails. Reaching that stage does not establish how well Gemini 4 reasons, codes or handles safety-critical requests.
7
10
Human feedback, reinforcement learning and specialized data can all be discussed as possible post-training methods, but the cited summit coverage does not establish which of them Google is using for Gemini 4 or how much each contributes. Nor does it provide measured improvements in reasoning, safety or reliability. Those details should not be inferred from the phrase “early post-training.”
13
19
Release early, then iterate
Kavukcuoglu said Google intends to release an early post-training output as soon as possible, with fast iterations to follow. The reporting does not specify a launch date or explain exactly how real-world feedback would be collected and incorporated into later versions.
13 His “much earlier” target expresses confidence in the schedule, not a commitment to a particular month.
7
Why the update matters
The remarks came during Kavukcuoglu’s first media appearance as DeepMind’s new day-to-day leader. Demis Hassabis stepped down as CEO in August to become DeepMind chairman and Alphabet chief scientist; he did not leave Google.
19
20
The Information frames Gemini 4 as a long-awaited flagship release amid competition with Anthropic and OpenAI.
19 Google has also continued releasing other Gemini models: a model-history source lists a Gemini 3.1 Pro preview in February 2026, while Google’s release notes describe Gemini 3.5 Flash as designed for speed and challenging tasks. Those facts complicate claims that Google has shipped no meaningful model updates or that Flash models are simply weak substitutes.
32
31
For now, the test is delivery rather than rhetoric. A release would let users assess Gemini 4’s reasoning, coding, reliability and cost against available alternatives. The summit update supplies no public results that settle those comparisons, so claims that Gemini 4 has restored Google’s position at the AI frontier are premature.
13
19