Google Confirms Gemini 4 Is in Post-Training and Coming 'As Soon as Possible'

Google DeepMind chief scientist Koray Kavukcuoglu confirmed this week that Gemini 4 has entered post-training — the phase where behavior is refined, safety tested, and guardrails implemented before public release.

Speaking at The Information's AI Agenda Live Summit, Kavukcuoglu said Google hopes to ship Gemini 4 "much earlier" than end of 2026, without providing a specific date. Internal testing is already using Gemini 4 to power Antigravity, Google's agentic coding environment.

Catching up in a crowded release window

The confirmation arrives days after Anthropic's Claude Opus 5.5 launch and OpenAI's GPT-6 Sol and Luna releases — intensifying perception that Google must prove frontier parity, not just ecosystem scale.

Kavukcuoglu pushed back on "fallen behind" narratives: "I have the utmost trust in the team" and "in my mind, it's a certainty that we are always gonna be at the frontier."

The Gemini 3.x detour

Google's last major flagship was Gemini 3 Pro in November 2025, followed by Gemini 3.1 Pro in February 2026. The company teased Gemini 3.5 Pro over the summer but never released it. Kavukcuoglu said Google "took a little bit of a step back" to focus on Flash models, including recent Gemini 3.8 Flash updates.

Current focus is squarely on Gemini 4 as the next flagship milestone.

Parallel Google moves this week

Gemini 4 timing intersects with other major Google AI infrastructure stories:

Project Suncatcher — an orbital AI data center test launching October 1
September 2026 spam update — SpamBrain refinements affecting AI Overviews enforcement
SAFA discussions — reported collaboration with OpenAI and Anthropic on frontier AI safety standards

Google is simultaneously racing on models, infrastructure, search policy, and industry governance.

What post-training implies

Post-training typically includes RLHF-style alignment, red-teaming, tool-use safety evaluation, and deployment gating. Given this week's OpenAI agent breach of an Australian government portal, safety testing for agentic capabilities will face heightened scrutiny — including Google's own agent products.

Developer implications

If you build on Gemini via Vertex AI or Google AI Studio, Gemini 4 will likely introduce new context windows, tool APIs, and pricing tiers. Planning for multi-model fallback — already best practice — becomes essential when flagship releases cluster in single weeks.

Google's public messaging is confident. The market's verdict will arrive when benchmarks, developer adoption, and real-world agent behavior meet the post-training promises.