Google Says Its Next Big AI Is Coming Very Soon
Google plans to launch Gemini 4 "much earlier" than the end of the year. The system is now being fine-tuned. The Verge
The timing was shared by Koray Kavukcuoglu, leader of Google's DeepMind division, in his first press interview in the role. He spoke to The Information, in comments described on Sept. 24, 2026. Google has not launched a major new model since the Gemini 3 series in November 2025.
Kavukcuoglu took over from former head Demis Hassabis, who left in August. He had led the teams that built Gemini and was asked to bring it to all Google products. He was Chief Technology Officer of Google DeepMind and Chief AI Architect before becoming SVP of Google DeepMind. He will now lead Gemini and report to CEO Sundar Pichai. CNBC
Kavukcuoglu said Google "took a little bit of a step back" to focus on Flash models, which are faster and less powerful, instead of releasing Gemini 3.5 Pro. The company chose speed and lower running costs over a middle-step upgrade.
He said Gemini 4 is in early post-training. That is the step after basic training when the model learns to give helpful and safe answers, like a new hire learning on the job. The Information It includes learning to follow instructions, follow rules, use tools, and pass repeated tests for real questions, long chats, and multi-step jobs.
The broader context here is how Google spaces out releases. Skipping a 3.5 Pro saved testing, safety and launch work while keeping Google's large systems running smoothly.
In my view, that choice looks practical. Top labs now compete on low cost and steady service as much as on high test scores, and Flash-type models serve much of daily business and consumer use.
Looking at what this means for app makers, timing and reach matter together. Kavukcuoglu leads Gemini across Google products, and a launch "much earlier" than year-end would leave teams little time to re-test search links, tool links, photo and voice inputs, and safety checks.
Worth flagging is that behavior can still change during fine-tuning, so early impressions should be treated as early only.
Looking ahead, if timing holds, developers get a new top model to try and new material to build smaller models from. That should help real products.
