Google DeepMind chief says Gemini 4 releasing much earlier than year-end

By Splitfeed EditorialLast updated Sep 26, 2026, 19:17 UTC

On September 24, 2026, Koray Kavukcuoglu, the recently appointed head of Google DeepMind, told The Information that Gemini 4 is currently in post-training and will be released much earlier than the end of the year. The announcement came three days after leaked Arena testing, first reported on September 21, surfaced the model with details suggesting it could deliver costs roughly five times lower than its predecessor and introduce new 3D capabilities. Gemini 4 is already running internally inside a coding tool called Antigravity. The timing is notable: Gemini 3.5 Pro was quietly discontinued ahead of the anticipated launch, and multiple senior DeepMind researchers have departed for OpenAI and Anthropic in the period leading up to this announcement.

The remarks carry weight beyond a simple product timeline update. Kavukcuoglu used the occasion to distance himself from the AGI-centered framing associated with his predecessor Demis Hassabis, calling that framing not the right conversation and placing emphasis instead on trustworthy agents. That is a meaningful rhetorical shift for a lab that had made AGI pursuit central to its public identity. The cost reduction figures from the leaked Arena tests, if accurate, would also matter commercially, since pricing has become a key competitive variable against OpenAI and Anthropic. The staff departures add a layer of uncertainty about whether Google can maintain the research depth needed to compete at the frontier.

The pro-AI camp has not yet published reactions to this story. Based on how that camp typically responds to major model releases, it would likely highlight the cost reduction and accelerated timeline as signs that AI capability and affordability are advancing faster than critics anticipated. The anti-AI camp has focused less on Gemini 4 specifically and more on the broader pattern of Google embedding Gemini across its products. Accounts including Kastor and Peter Flax have criticized the model for appearing in services like Google Maps without user request and for producing directional claims that fall apart under scrutiny. Barry Schwartz pointed to Gemini Notebook content filling Google Search as a preparation and quality failure. The middle ground camp has also not reacted publicly, but would typically ask whether the faster release schedule reflects genuine readiness or competitive pressure overriding caution, particularly given the hallucination complaints already attached to earlier versions.

The clearest near-term signal will be the actual release date and whether Google publishes transparent benchmarks alongside it, including on reliability and factual accuracy. Any formal disclosure about the security testing incident referenced in pre-release reporting would also shift the debate significantly.