Gemini 4 Enters Post-Training — Google's Concrete Signal in the Three-Way AI Race

Google's Gemini 4 has entered post-training, according to Koray Kavukcuoglu, the newly elevated Senior Vice President of Google DeepMind, speaking at The Information's AI Agenda Live Summit on September 23–24, 2026. It is the first public signal of where Google's next frontier model sits in the development pipeline — and the clearest read yet on where Google stands in the three-way race with OpenAI and Anthropic.

Google has not officially announced Gemini 4. The post-training milestone comes from Kavukcuoglu's remarks at a summit, not from a Google blog post, model card, or press release. That distinction matters: what is known is the pipeline stage and the executive signal; what is not known is a release date, a capability profile, or any benchmark.

What post-training means here

Post-training is the phase that follows a model's initial pre-training run — the tuning, alignment, and refinement work that turns a raw trained checkpoint into something that can be shipped. A model in post-training is not a model ready to ship the next day, but it is past the most compute-heavy and least predictable stage of development. Pre-training is the long, expensive, capacity-constrained run; post-training is where teams squeeze capability, safety, and usability out of what the pre-training run produced.

For a frontier lab, announcing or confirming a post-training entry is more concrete than saying a model is "in development" — it means the pre-training run is done and the team is now working the model toward a shippable state.

The timeline so far

Kavukcuoglu confirmed that Gemini 4's pre-training run began on July 21, 2026. As of late September, that puts the pre-training phase at roughly two months, followed by the post-training phase now underway. The elapsed time is a data point, not a benchmark — different labs and different model generations take different amounts of pre-training time, and the elapsed clock says less than the team's own progress markers do.

The public comparison point is Gemini 3. Google's Gemini 3.8 family is already live as of this writing, with the Gemini 3.8 Flash TTS voice models shipping on September 26 (AIPress, Post 49). That means Google is simultaneously running its current-generation models and pushing its next-generation model through post-training — a normal pattern at the frontier, but one that signals the Gemini 3 generation is not the end of Google's roadmap.

Where Google sits in the three-way race

The three frontier labs are moving on different clocks right now. A heat check as of late September 2026:

OpenAI — GPT-6 Astra shipped September 3. GPT-6 Sol and GPT-6 Luna followed on September 22. A GPT-6 Cyber cybersecurity preview is expected at DevDay on September 29. That is four model moves in under four weeks, including a security-specific model and a reportedly first-of-its-kind deployment product. OpenAI's cadence is the most active of the three right now.

Anthropic — Claude Fable 5.1 shipped September 20. Claude Opus 5.5 followed on September 22, landing in GitHub Copilot at 40% less cost than the equivalent prior flagship. Anthropic has also been running a heavier governance and safety narrative — the September 2026 threat report (AIPress, Post 6), the Accenture embedded-evaluation partnership (Post 48), and the public clash with Microsoft over training Claude to reason about its own moral status (Post 47).

Google — Gemini 3.8 is live and shipping feature additions (the Flash TTS models on September 26). Gemini 4 is in post-training. No Gemini 4 ship date is public. Google's posture is quieter than OpenAI's headline cadence and less safety-narrative-heavy than Anthropic's, but the post-training entry is a concrete marker that the next model is not speculative.

The framing that matters for readers: Google is not stalled. The post-training entry means the next Gemini is past the hardest compute stage and into the finishing work. How fast that finishing work moves, and what Google chooses to ship first, is the open question.

What Kavukcuoglu's signal is worth

Kavukcuoglu was elevated to Senior Vice President of Google DeepMind on August 12, 2026 — his first public appearance in the role was this summit, and he used it to confirm the post-training milestone and to say he wants Gemini 4 out "well before the calendar turns." That is a stated goal, not a confirmed launch date. The Information's reporting and downstream tech coverage from tech-insider.org and shattered.io have picked the same thread, all attributing the timeline to Kavukcuoglu's summit remarks.

A senior executive saying he wants a model out before year-end is a directional signal, not a commitment. It tells readers that Google's internal target is aggressive; it does not tell readers what will actually ship, when, or with what capabilities.

The comparison that matters

The most useful frame for readers following the three-way race is not "who has the best model today" — that shifts with every release — but "who is moving, and how"::

• OpenAI is moving hardest right now in terms of release cadence and product breadth (model + deployment product).
• Anthropic is moving on capability and on a safety/governance narrative simultaneously.
• Google is quieter on the public roadmap but has Gemini 4 past pre-training and into post-training, with an executive publicly targeting an end-of-year or earlier window.

A post-training entry does not mean Gemini 4 will beat GPT-6 Astra or Claude Opus 5.5 on any given benchmark. It means Google is not standing still while the other two labs ship. That is the headline for readers who track the race rather than any single model.

The honest caveat

Gemini 4 has not been officially announced by Google. There is no model card, no benchmark, no release date, and no capability list. Everything in this post traces back to Kavukcuoglu's remarks at a summit, as reported by Yahoo Finance and downstream tech outlets. Treat the timeline as a reported target, not a schedule, and treat any comparison to OpenAI or Anthropic's current models as a positioning read rather than a capability read — there is no public Gemini 4 data to compare against yet.

Related AIPress coverage

Gemini 3.8 Live Review: Google Takes #1 on Speech-to-Speech Quality With 82.6 Score — the current-generation Gemini that Gemini 4 will succeed: https://aipress.blog/post/gemini-3-8-live-review-google-takes-1-on-speech-to-speech-quality-with-82-6-score

Google Launches Gemini 3.8 Flash TTS — Two Voice Models Built for Scripting and Volume — the most recent Gemini shipping news, September 26: https://aipress.blog/post/google-launches-gemini-38-flash-tts-two-voice-models

GPT-6 Astra vs Claude Fable 5.1 vs Gemini 3.8: The September 2026 Benchmark Round-Up — the three-way comparison baseline that Gemini 4 will eventually be measured against: https://aipress.blog/post/gpt-6-astra-vs-claude-fable-5-1-vs-gemini-3-8-the-september-2026-benchmark-round-up

Published September 27, 2026. Sources: Yahoo Finance (September 24, reporting on Kavukcuoglu's remarks at The Information's AI Agenda Live Summit, September 23–24, 2026); tech-insider.org; shattered.io. Google has not officially announced Gemini 4. The post-training milestone, the July 21 pre-training start date, and the end-of-year target are all attributed to Kavukcuoglu's summit remarks, not to a Google announcement.