Google's Gemini 4 Enters Post-Training, With No Release Date Yet

Google DeepMind has moved Gemini 4 into early post-training, according to remarks from its senior vice president, Koray Kavukcuoglu, at a September AI conference. No release date has been set. The most detailed reporting found for this article, from The Information and The Verge, is dated 24 to 26 September 2026, so the news is a few weeks old, and no newer official update turned up.

Kavukcuoglu was promoted to senior vice president on 12 August. He made his first public appearance in the role at The Information's AI Agenda Live Summit on 23 and 24 September, according to a Yahoo Finance write-up of the event. He said Google intends to release an early version of the post-training model as soon as possible, but the article gives no date for that release.

What the Gemini 4 timeline means for Google.

Google's last flagship family, Gemini 3, launched in November 2025. Since then, OpenAI has released GPT-6 and Anthropic has released its Mythos series. The Northeast Times, citing The Verge, says both outperform Gemini 3. Sundar Pichai had promised a major Gemini 3.5 Pro update by June, and as of late September that model had not shipped publicly. Kavukcuoglu said the company has shifted attention to faster, less powerful Flash models instead.

A benchmark cited in the Yahoo Finance article puts Google's current frontier model, Gemini 3.6 Flash, at 34.0 on the Intelligence Index v4.3.2, against 57.6 for Anthropic's Opus 5.5. The article does not say who produced those scores, so the gap should be read with caution. The market also reacted: Alphabet shares fell nearly 5% in the days before the interview, amid concern about Meta's new Muse AI hardware, according to a startupfortune.com report that the Northeast Times cited.

On the technical side, Google said it began its largest pre-training run for Gemini 4 on 21 July 2026, according to a Google blog post cited by Yahoo Finance. Kavukcuoglu described it as the company's most ambitious pre-training run to date. Google says the model is considerably larger than its predecessors, but it has published no parameter count, architecture or pricing.

Several questions remain open. Google has not given a release date, benchmark results or pricing. Whether a compressed post-training schedule leaves enough time for safety evaluation is a question the reporting raises without answering. Because the coverage traces back to a small number of outlets and Google has not made a formal announcement about this stage, this article is shorter than a full story would be.

For readers tracking this, the signal to watch is an official Google post that names the model and gives a date. Until then, anything beyond the sourced claims above is speculation. Developers who compare model outputs can tidy the JSON responses with the free formatter at /en/json-formatter.