Koray Kavukcuoglu takes over DeepMind as Demis Hassabis steps back
On August 12, 2026 Google announced that Koray Kavukcuoglu, DeepMind’s former CTO, is replacing Demis Hassabis as head of the AI unit, with a mandate refocused on the frontier and on code. The move is a deliberate pivot from research toward product execution to catch OpenAI and Anthropic.
August 12, 2026. Google announced that Koray Kavukcuoglu is taking over DeepMind, replacing co-founder and CEO Demis Hassabis, who becomes the unit’s chair. The former CTO and chief AI architect will now report directly to Sundar Pichai. The same day, Alphabet shares slipped — a sign that the market reads this reshuffle as an admission of a frontier gap Google wants to close fast.
The subtext is plain: Google has not shipped a frontier model since early 2026, while OpenAI and Anthropic have kept releasing widely praised systems. Changing captains means choosing execution over research.
This is no simple game of musical chairs. It comes after months of executive departures and a delayed flagship model, at a moment when the race has tightened around a single issue: shipping frontier models that are competitive at code. The choice of Kavukcuoglu says one thing plainly — Google wants to win that race, and fast.
The diagnosis: Google is falling behind at the frontier
Since ChatGPT in 2022, OpenAI has set the pace. Google briefly reclaimed the lead in November 2025 with Gemini 3, a model hailed for having “moved the frontier forward”. But in 2026 the dynamic flipped: Anthropic launched Mythos, OpenAI shipped GPT-5.6, and Google, after Gemini 3.1 Pro in February, has not fielded a model that challenges at the top.
The gap concentrates on one specific battleground: code. Analysts are blunt: Google is “miles behind” OpenAI and Anthropic on coding, says Malik Ahmed Khan of Morningstar, because it prioritized monetizable areas like search and multimodal over the LLM’s first killer use case.
The most tangible signal is the delay of Gemini 3.5 Pro: in July, Bloomberg reported Google had pushed the release back to try to improve its performance, particularly in programming. When a giant delays its flagship model to make it better at code, the pressure has become concrete.
Why code became the battleground
Google’s gap is measured first in code for a simple reason: that is where generative AI pays. Coding assistants and coding agents are the first use case enterprises pay for at scale — a market OpenAI and Anthropic now dominate with tools tied directly to their frontier models. A model that leads on code pulls everything else with it: developers adopt the API, the company buys the seats, and the vendor collects the usage data that feeds the next model.
That is the virtuous cycle Google is trying to join with Code Strike, the internal team assembled to strengthen programming capabilities. The delay of Gemini 3.5 Pro follows the same logic: shipping a model still trailing on code would be worse than waiting, because a disappointing launch on that front translates immediately into lost developer market share. In 2026 the frontier race is no longer about general benchmarks — it is about the compiler output, the tests, and the diffs an agent produces correctly.
Code also has a property other uses lack: it is verifiable. A model that writes a program can be judged at execution — the test passes or it fails — where a conversational answer is judged on approximation. That objective ground is why labs see coding as the shortest path to a reasoning-capable AGI.
Who Koray Kavukcuoglu is
Kavukcuoglu is no newcomer: he joined DeepMind in 2012, before the Google acquisition in 2014. He is part of the old guard, having already served as the unit’s CTO and the parent company’s chief AI architect. Over the past year he had been absorbing Hassabis’s responsibilities, directing model development and presenting major Gemini releases.
The contrast with Hassabis is deliberate. Where the co-founder “was much more interested in building AGI that was beyond LLMs”, in Malik Ahmed Khan’s words, Kavukcuoglu is described as an execution profile. Ray Wang of Constellation Research sums it up: his promotion “shows that Google is prioritizing execution over deep research”. The expected consequence is a faster release cadence, more experimentation, and a product roadmap.
What changes structurally
Beyond the people, the org chart tells the story of the pivot. Kavukcuoglu will oversee Gemini model development, frontier AI research, and the Gemini app and developer teams — all while reporting directly to Sundar Pichai.
That concentration is a structural choice, not cosmetics. “Models, the app, and the developer teams under one executive is how a company organizes a product group rather than a lab,” says Brian Hopkins of Forrester. It is the exact inverse of the machine Hassabis embodied: research fundamentally aimed at AGI, with work in health (Isomorphic Labs) and world models.
The change also has a geography. Kavukcuoglu, based in Mountain View, symbolizes the shifting center of gravity of Google’s AI away from London, where Hassabis had stayed. “Google is quietly consolidating its AI leadership out of London,” notes Nick Patience of the Futurum Group — a detail that matters for the UK AI ecosystem, of which DeepMind was long the flagship.
The bet: execute, and fast
Kavukcuoglu’s implicit roadmap comes down to two missions. First, ship Gemini 3.5 Pro, then “prove it wasn’t a one-off by maintaining a predictable release cadence”, as Nick Patience puts it. Second, rebuild the coding and pretraining expertise that walked out the door — an allusion to the executive departures that have marked Google in recent months.
But one should resist the picture of a general crisis. “It’s a big team. It’s several thousand people on the Gemini team,” cautions Sebastian Mallaby, author of a Hassabis biography. Above all, the monetization is healthy: nearly 90% of Fortune 100 companies use Gemini Enterprise, and Google’s AI runs on Android phones and inside Apple Intelligence. Google sells what it has extremely well; it is the top of the range it is missing.
The risk of the pivot is symmetric: by refocusing DeepMind on product, Google could regain ground at the frontier — or dilute exactly the long-horizon research culture that produced AlphaGo and then Gemini 3. That is the classic dilemma of a lab turned into a product group.
Verdict
Koray Kavukcuoglu’s appointment is a deliberate strategic admission: Google is reorganizing to catch OpenAI and Anthropic at the frontier, especially on code, with an execution profile rather than a research one. It is a bet that the gap is a problem of focus and cadence rather than raw talent — a bet that a different captain can change the outcome.
The recommendation is conditional. If you build on the Gemini ecosystem, expect a faster release cadence and more product-oriented APIs — Kavukcuoglu’s mandate is precisely to bring models closer to developers. If you are betting on who wins the frontier race, watch the next quarter: a Gemini 3.5 Pro that is competitive at coding would make Google a credible contender again; another delay would confirm the problem was never just about the captain. And if you are an enterprise Gemini customer, you are the healthy part of the equation: monetization is strong, and the product refocus can only strengthen the support and roadmap you depend on.