Gemini 4 nears launch, new DeepMind chief says

Google is closing in on its next flagship model after a long dry spell. Koray Kavukcuoglu, who took over DeepMind after Demis Hassabis departed in August, told The Information that Gemini 4 is in final polishing and slated for release "well before the end of the year." It is the first concrete timeline the new leadership has offered after months of silence.

Since the Gemini 3 series arrived in November 2025, Google has not shipped a new flagship. In that window OpenAI released GPT-6 and Anthropic the Mythos series, both of which leapfrog Gemini 3 on standard benchmarks and open a significant lead. At I/O in May, Sundar Pichai promised a major Gemini 3.5 Pro update by June; the model never appeared.

Kavukcuoglu acknowledged the company "took a small step back" to prioritize Flash models — faster, lighter variants — over finishing 3.5 Pro. The decision reflects a shifted priority: a series of lighter models ready for immediate production instead of a single heavy model. The move kept Google present in the market but left the top of the pyramid exposed.

The plan, Kavukcuoglu said, is to ship a "post-training product as early as possible because we see the results and we're excited." He added the company will continue "rapid iterations" from here. The phrasing suggests Gemini 4 will arrive as an initial version improved through continuous updates — a pattern that resembles competitors' release strategies more than Google's traditional approach.

The statement comes two months after Hassabis left following an announcement of "major progress" on Gemini 4. Kavukcuoglu, formerly his deputy and the technical architect behind many DeepMind breakthroughs, now gets his first public turn. The question is whether his style — fewer grand pronouncements, more measured engineering — will translate into a consistent release cadence.

The promise to "always be at the frontier" sounds good in an interview, but the real test is performance: whether Gemini 4 narrows the gap with GPT-6 and Mythos on coding, complex reasoning, and long context windows. Until the weights ship and the community can run independent evaluations, this is a statement of intent, not proof.