Google DeepMind Confirms Gemini 4 in Post-Training, Eyes Launch Before Year-End

Speaking at The Information's AI Agenda Live Summit on September 23, new Google DeepMind chief Koray Kavukcuoglu confirmed that Gemini 4 has moved into post-training — the phase where a model's behavior is refined and safety guardrails are installed before public release. He set an aggressive expectation: an early version launching well before the end of 2026.

The subtext is competitive urgency. Google has not shipped a new flagship since Gemini 3 in November 2025, an unusually long gap by 2026 standards, during which OpenAI's GPT-6 family and Anthropic's Mythos series both launched and, by most benchmark accounts, pulled ahead. Kavukcuoglu's comments read as reassurance to a developer base that has watched rivals iterate faster.

There are concrete signals of progress: internal testing of Gemini 4 is already powering Google's Antigravity coding tool, and Google Cloud separately made Gemini 3.8 Live with Live Avatar generally available in Gemini Enterprise this week — evidence the pipeline is shipping, even if the flagship isn't.

Skepticism runs high. Developers on X and in the source reporting noted Google's stalled cadence since Gemini 3 and questioned whether a rushed 'early version' can close a benchmark gap against GPT-6 and Mythos. There is also unresolved reputational baggage: community chatter this week flagged concerns over Google's four-month delay in disclosing a Gemini model breach and internal worry that moving the AI-responsibility team could compromise independent safety research. The launch timeline vow will be judged against whether the shipped model actually competes, not merely arrives.