Google Targets an Early Gemini 4 Launch as Testing Continues
Google aims to launch Gemini 4 "much earlier" than the end of the year, with the model now in its refinement stage. The Verge
The timeline came from Koray Kavukcuoglu, leader of Google's DeepMind division, in his first media appearance in that role. He spoke in an interview with The Information, described on Sept. 24, 2026. Google has not released a new flagship AI model since the Gemini 3 series in November 2025.
Kavukcuoglu succeeded former DeepMind head Demis Hassabis, who stepped down in August. He had managed the teams building the Gemini models and was tasked with bringing Gemini to all Google products. He served as Chief Technology Officer of Google DeepMind and as Chief AI Architect before becoming SVP of Google DeepMind. In the role he will oversee Gemini and report directly to Google CEO Sundar Pichai. CNBC
Kavukcuoglu said Google "took a little bit of a step back" to focus on faster, less powerful Flash models instead of releasing Gemini 3.5 Pro. The choice favored response speed and lower running costs over a mid-step Pro upgrade.
On technical status, Kavukcuoglu placed Gemini 4 in the early days of post-training, the phase where a base model is tuned to follow instructions reliably. The Information That work includes instruction tuning, alignment, tool use and repeated testing, which show how well the raw model handles real prompts, long conversations and multi-step tasks.
The broader context here is release pacing. Skipping a 3.5 Pro to ship Flash variants saves testing, safety and deployment work while keeping Google's serving systems efficient.
In my view, that tradeoff looks practical rather than cautious. Leading labs now compete on cost per token and steady output as much as on top test scores, and Flash-type models handle much of everyday business and consumer use.
Looking at what this means for builders, timing matters alongside reach. Kavukcuoglu's mandate covers Gemini across Google products, and a flagship arriving "much earlier" than year-end would give teams a short window to re-test search links, function calling, image and voice inputs, and safety controls against the new model.
Worth flagging here is that early post-training behavior can still shift before release, so current impressions should be treated as provisional.
Looking ahead, if the schedule holds, developers get a new top model to test and fresh options for building smaller models from it. That pairing should help real production work.
