On 5 August 2026 Sundar Pichai announced that Demis Hassabis becomes chair of Google DeepMind and chief scientist of Alphabet, while retaining leadership of Isomorphic Labs. Koray Kavukcuoglu becomes senior vice president of Google DeepMind, overseeing Gemini model development, frontier AI research, and the Gemini app and developer teams, reporting to Pichai, and keeps his chief AI architect title.
In the same message, Google confirmed that Jeff Dean is leaving after 27 years to start a public benefit corporation with Sanjay Ghemawat, with Google as founding investor.
Reorganisations are usually less interesting than they look. This one is worth reading closely, because of what Google’s product line has and has not done this year.
The gap the reorganisation points at
Google’s Flash tier has moved fast. Gemini 3.6 Flash shipped on 21 July 2026 with text, image, video, audio, and PDF input, built in computer use, and a reported 17 percent reduction in output tokens against its predecessor. It is the broadest input coverage of any current model and it is genuinely good.
The Pro tier has not moved. Gemini 3.1 Pro launched on 19 February 2026 and, nearly six months later, still ships under a preview model identifier. Google has confirmed that Gemini 3.5 Pro remains in testing and that Gemini 4 is only in pre training.
So in a year when OpenAI shipped an entire three model family and cut its prices twice, and Anthropic shipped Fable 5, Sonnet 5, and Opus 5 inside seven weeks, Google’s frontier tier has stood still while its workhorse tier has been excellent. That is not a research problem. Google has as much research depth as anyone. It is a problem of converting research into shipped, supported, generally available product on a competitive cadence.
Why separating the two roles makes sense
Research and shipping want opposite things from a leader’s calendar. Research needs long horizons, tolerance for dead ends, and protection from quarterly pressure. Shipping needs someone who will freeze a model, cut the last three features, and put a stable identifier on it.
Asking one person to hold both is how you get a preview label that lasts six months. Splitting them gives Hassabis the long horizon and Alphabet wide scientific remit that suits him, and gives Kavukcuoglu a single accountable job: get Gemini out of the door.
Kavukcuoglu reporting to Pichai rather than to Hassabis is the detail that matters most. It means the shipping decision no longer has to route through the research organisation.
The Jeff Dean departure is the bigger story
Dean spent 27 years at Google and his fingerprints are on much of the infrastructure the modern industry runs on. That Google is a founding investor in the public benefit corporation he is starting with Sanjay Ghemawat suggests an amicable, planned exit rather than a rupture.
It is still a substantial loss of institutional knowledge, and it lands in the same announcement as a leadership restructure. Reading them together, this looks less like a routine reshuffle and more like a deliberate generational handover of the AI organisation.
What to watch
One thing, really. Whether Gemini 3.1 Pro loses the preview label, or a successor ships as generally available, before the end of 2026. If it does, the reorganisation worked. If Google is still shipping a preview flagship in six months while its Flash tier keeps improving, the problem was never the org chart.