In Brief: Demis Hassabis has handed day-to-day management of Google DeepMind to Koray Kavukcuoglu and become chair of the lab and chief scientist of Alphabet. As senior vice president, Kavukcuoglu is now responsible for Gemini development, frontier research, the Gemini app, and developer teams. After 27 years at Google, Jeff Dean is launching an independent science company with Sanjay Ghemawat. Google is not winding down Gemini: the company has officially confirmed work on Gemini 4, but has not announced a release date or specs.
This article is intended for CTOs, AI/product leaders, and engineering executives. We cover the confirmed appointments and what they mean organizationally; rumors about internal conflicts, investment forecasts, and unconfirmed Gemini 4 specs are not discussed.
Contents
- What happened at Google DeepMind
- Who is responsible for what now
- What is known about Gemini 4
- Why Google is separating science and delivery
- What Jeff Dean's departure means
- What changes for Gemini customers
- The 3-layer AI team model
- AI team audit: 7 questions
- Frequently asked questions
- Bottom line
What happened at Google DeepMind
On August 5, 2026, Google announced the biggest leadership overhaul of its AI division since the merger of DeepMind and Google Brain in 2023. In the official announcement Sundar Pichai and Demis Hassabis outlined three changes.
First: Hassabis is no longer running the daily operations of Google DeepMind. He became Chair of Google DeepMind and Chief Scientist of Alphabet. In this role, he will focus on long-term AGI strategy, science, and global issues while continuing to lead Isomorphic Labs.
Second: Koray Kavukcuoglu became senior vice president of Google DeepMind and reports directly to Sundar Pichai. He retains the role of Chief AI Architect at Google and gets end-to-end responsibility for Gemini models, frontier AI research, the Gemini app, and developer teams.
Third: Google chief scientist Jeff Dean and Google Senior Fellow Sanjay Ghemawat are creating an independent public benefit corporation to accelerate research in machine learning, science, and engineering. Google is acting as a founding investor and cloud partner.
This is not a formal replacement of one CEO with another. The new structure splits responsibility between Alphabet's long-term science agenda and Gemini's operational delivery.
Who is responsible for what now
| Person or layer | Was | Now | Practical meaning |
|---|---|---|---|
| Demis Hassabis | CEO of Google DeepMind | Chair GDM + Chief Scientist of Alphabet | AGI, science strategy, external influence, Isomorphic Labs |
| Koray Kavukcuoglu | CTO GDM + Chief AI Architect Google | SVP Google DeepMind + Chief AI Architect | single accountability for models, research, the app, and developer delivery |
| Sundar Pichai | CEO of Google and Alphabet | direct manager of Kavukcuoglu | shorter link between DeepMind and Alphabet's business |
| Jeff Dean and Sanjay Ghemawat | Chief Scientist / Google Senior Fellow | independent science company | research outside the constraints of a public corporation, but in partnership with Google |
| owner and infrastructure base of GDM | investor and cloud partner for the new project | retains a connection to the separated research layer |
Important detail: Kavukcuoglu did not receive the title of CEO of Google DeepMind, but rather the SVP position. That strengthens the direct management link to Alphabet. At the same time, the official scope is broad: model development, frontier research, the Gemini app, and developer tools are all under one operational leader.
What is known about Gemini 4
Only the fact that Gemini 4is being developed has been officially confirmed. Hassabis said the models are “in good hands” and mentioned progress on new models, including Gemini 4. Google did not disclose a release date, context size, pricing, benchmark results, or a list of available APIs.
So the correct wording as of August 6, 2026, is:
Gemini 4 is in development, but the release date and technical specs have not been publicly announced. Any exact timelines, pricing, and benchmark scores outside the official announcement should be treated as unconfirmed.
Google also shared several numerical signals that show the scale of the current ecosystem:
- the Gemini app reached more than 950 million monthly active users;
- Gemma models surpassed 900 million downloads;
- Kavukcuoglu has worked at DeepMind for 13 years;
- Jeff Dean is ending his 27-year stint at Google.
All four metrics are measured according to Google, but the first two are the company's own data, not an independent audit. These figures explain why Kavukcuoglu's new role combines not only research, but also a consumer-scale app with a developer ecosystem: mistakes in the release train now affect hundreds of millions of users.
Why Google is separating science and delivery
In 2023, the logic was the opposite: Google merged Google Brain and DeepMind to bring researchers, compute, and model development under one division. At the time, the company wrotethat the merger should accelerate the creation of more capable and responsible AI systems. Hassabis became CEO, and Dean became Google’s chief scientist.
In 2026, Google is not splitting the teams apart again. It is separating the type of decisions:
- Hassabis gets time for questions with a years-long horizon: AGI, AI for science, safety, and engagement with governments and society.
- Kavukcuoglu is responsible for decisions with a weeks-and-quarters horizon: model roadmap, evals, releases, the app, APIs, and execution speed.
- Pichai becomes the direct escalation point for all of Google DeepMind.
This kind of split removes calendar conflict. A scientific leader may consider it right to explore a new class of architectures even if it does not improve the product in the current quarter. A product leader must manage latency, inference cost, safety, API backward compatibility, and the release date. When both sets of decisions are tied to one person, the urgent almost always crowds out the important — or vice versa.
Axios connects the restructuring with model delays, researcher departures, and pressure from OpenAI and Anthropic. In its official statement, Google does not name these factors as the cause and emphasizes strong demand for Gemini. So a “reorganization because of failure” is an outside interpretation, not an established fact.
What Jeff Dean’s departure means
Jeff Dean and Sanjay Ghemawat shaped Google’s infrastructure for decades — from search systems to distributed machine learning. Their new project is supposed to accelerate discovery loops in ML, science, and engineering. According to Axios reports, Oriol Vinyals and Quoc Le are also joining them; Google will be an investor and cloud partner.
The financial terms have not been officially disclosed. Reports about the size of the round should be treated as preliminary until the company or investors publish documents.
The organizational meaning matters more than the amount: frontier researchers are looking for an environment where they can optimize not quarterly revenue, but the speed of scientific discovery. Google is trying not to lose connection to that layer entirely — it is supporting a separate organization with capital, compute, and research partnership.
For the AI market, this reinforces the trend toward labs built around the experimental cycle: an agent forms a hypothesis, plans an experiment, launches a simulation or lab process, analyzes the result, and chooses the next experiment. Such long-lived processes require memory, checks, and observability — we covered the architecture of these systems in the article “Asynchronous AI Agents: Architecture and Memory”.
What changes for Gemini customers
For Gemini and Google Cloud users, no immediate API or pricing changes have been announced. A personnel news item cannot be used to conclude that a model, contract, or endpoint will automatically change.
But enterprise buyers should watch four practical signals:
| Signal | What to check | Customer decision |
|---|---|---|
| Release cadence | frequency and predictability of model updates | keep an abstraction layer and canary tests |
| API stability | deprecations, rate limits, compatibility | do not tie a critical process to a single model ID |
| Eval transparency | model cards, safety notes, reproducible evals | update your own benchmark before migration |
| Product integration | Gemini app, Workspace, Cloud, AI Mode | compare the value of integration with the risk of vendor lock-in |
Bottom line: a change in leadership is not a signal to urgently switch vendors, but it is a reason to check how reversible your integration is. A good AI architecture allows the model to be swapped without rewriting business logic, data, access control, or logging. Practical criteria for choosing a vendor and platform are collected in the article “How to Choose an LLM Integrator”.
The “3 Layers of an AI Team” model
It is useful to translate Google’s restructuring into a model that fits companies of any size. AI Dawn calls it “3 Layers of an AI Team”. This is an analytical framework, not a Google term.
Layer 1. Frontier and research
The goal is to look for opportunities that are not yet in the roadmap: new methods, types of agents, training approaches, scientific applications. The metrics are the quality of hypotheses, reproducibility, the team’s learning rate, and the number of decisions that move into engineering. The timeframe is quarters and years.
Layer 2. Model delivery
The goal is to turn capability into a reliable service. This is where evals, red teaming, latency, cost per task, model routing, rollout, and rollback live. The metrics are task success rate, incidents, cost, speed, and the share of safe rollbacks. The timeframe is weeks and months.
Layer 3. Applications and adoption
The goal is to embed AI into a specific process and deliver business impact. The team is responsible for UX, human-in-the-loop, integrations, user training, and unit economics. The metrics are cycle time, output quality, adoption, and operating cost. The timeframe is days and quarters.
In a startup, all three layers may consist of four people. What needs to be separated is not headcount, but the backlog, metrics, and final decision rights. Research should not promise a production date; delivery should not present an experiment as a core strategy; product should not bypass evals for the sake of a pretty demo.
AI Team Audit: 7 Questions
- Who has the authority to say “the model is ready for production,” and based on which evals?
- Is the research backlog separated from customer commitments for the next 90 days?
- Is there a single owner of model delivery — cost, latency, safety, and rollback at the same time?
- Can the application team replace the model without rewriting workflows and access policies?
- What percentage of AI features goes through a canary rollout to 10% of users?
- Where is the decision log stored: why the model was updated, delayed, or rolled back?
- What happens if the lead researcher or AI architect leaves tomorrow?
If you do not have a concrete answer to three or more of these questions, the problem is not choosing Gemini, Claude, or GPT. The company has not yet defined an operating model for AI.
Frequently Asked Questions
Did Demis Hassabis leave Google DeepMind?
No. He stopped running day-to-day operations, but remained at Google DeepMind as chair and became chief scientist at Alphabet. He also continues to lead Isomorphic Labs and advise model and research teams.
Who is leading Google DeepMind now?
Kavukcuoglu is handling day-to-day operations as senior vice president of Google DeepMind. He reports to Sundar Pichai and is responsible for Gemini model development, frontier AI research, the Gemini app, and developer teams.
When will Gemini 4 be released?
No date has been announced. The official material confirms work on Gemini 4, but does not disclose the schedule, pricing, API, or benchmark results.
Why did Jeff Dean leave?
Officially, Dean wants to try a new format and, together with Sanjay Ghemawat, is creating an independent public benefit corporation for ML, science, and engineering. The company has not published more specific personal reasons.
Is Google DeepMind falling behind OpenAI and Anthropic?
Axios writes about delays and competitive pressure. Google emphasizes strong demand and Gemini’s scale. The comparison should be based on specific models on your own task set; a personnel restructuring by itself is not a benchmark.
Should businesses stop using Gemini?
No. The announcement does not automatically change the API or contracts. For businesses, it is more useful to review vendor portability, your own evals, and a rollback plan than to react to a personnel update with a rushed migration.
Conclusion
Google DeepMind did more than just change leadership. Alphabet split two timelines: Hassabis is responsible for the long horizon of AGI and science, while Kavukcuoglu is responsible for the pace and consistency of Gemini delivery. At the same time, Jeff Dean's departure shows that independent AI-for-science labs are becoming a separate magnet for talent and capital.
For business, the lesson is practical: separate research, model delivery, and applications at least at the level of responsibility. Assign an owner for production evals, keep the model layer replaceable, and document decisions. Then a leadership change at a supplier will remain a vendor review issue, not an outage for your product.