Signals Inbox·August 30, 2026·Frontier AI

Who is Thibault Sottiaux, product lead at OpenAI?

Thibault Sottiaux is OpenAI’s Head of Product & Platform, running the core product layer across ChatGPT, Codex, the API, agents and enterprise. His unusually fast rise from DeepMind engineer to one of OpenAI’s most influential product builders says a lot about where the company thinks AI products are going next.

Q1What does Thibault Sottiaux actually run at OpenAI today?

Thibault Sottiaux currently runs OpenAI’s core product organization, covering ChatGPT, Codex, the API, agent infrastructure and enterprise products.

That is broader than the shorthand “product lead.” OpenAI’s current Build Week page lists Sottiaux as Head of Product & Platform. In a recent TechCrunch interview, he described his own scope even more clearly: all core products, including the API, agent infrastructure, enterprise, ChatGPT and Codex.

He reports to Greg Brockman, who oversees OpenAI’s broader product strategy. Sam Altman remains CEO, so Sottiaux is not OpenAI’s ultimate product decision-maker. His unusually powerful position comes from owning the layer where OpenAI’s models become products people and companies actually use.

That scope has expanded very quickly. Sottiaux joined OpenAI in 2024, became closely associated with Codex, and now has responsibility for products ranging from a consumer chatbot used at huge scale to developer infrastructure and autonomous agents.

Q2How did Thibault Sottiaux go from DeepMind engineer to OpenAI product boss?

Thibault Sottiaux reached the top of OpenAI’s product organization through engineering, AI infrastructure and model development rather than the usual product-management career path.

Sottiaux grew up in Belgium and studied applied mathematics and computer science at Université catholique de Louvain. He joined Google in 2015, initially working as a software engineer before moving to DeepMind.

At DeepMind, his work moved progressively closer to the machinery behind frontier AI. He worked on machine-learning infrastructure, model serving, research workflows and eventually human-data operations for Gemini. The Gemini 1.5 technical report names him as one of two leads for human data.

He joined OpenAI in 2024 and initially returned to familiar territory: tools that helped OpenAI researchers work more effectively. Within months, that work moved toward coding agents and eventually Codex.

The unusual part of his career is the speed of the final step. He spent roughly nine years inside Google and DeepMind before joining OpenAI, then went from internal research tooling to running OpenAI’s core products in roughly two years.

Q3What did Thibault Sottiaux really do at Google DeepMind?

Thibault Sottiaux was mainly an AI infrastructure and data leader at DeepMind, with documented work spanning reinforcement learning, large language models and Gemini.

The AlphaGo connection needs some precision. WIRED reports that Sottiaux helped build infrastructure and tools used by DeepMind researchers working on systems such as AlphaGo. That puts him around the technical foundation that supported the research rather than among the small group best known for AlphaGo’s core algorithms.

His published work gives us a clearer picture. Sottiaux co-authored Reverb, DeepMind’s distributed experience-replay system for reinforcement learning. The Gopher paper credits him with model serving, meaning he worked on the systems required to actually run one of DeepMind’s early large language models at scale.

Gemini gives us the strongest evidence of senior responsibility. The Gemini 1.5 technical report explicitly lists Sottiaux as Lead, Human Data alongside Amelia Glaese. He also appears among the contributors to Imagen 3.

Across those projects, the recurring theme is practical AI systems: getting research infrastructure, data and models to work together reliably at large scale. That background explains a lot about the way he now approaches agents at OpenAI.

Q4Why did ChatGPT make Thibault Sottiaux want to leave DeepMind?

ChatGPT appears to have pushed Thibault Sottiaux toward OpenAI because he was frustrated by how slowly powerful AI research became something ordinary people could actually use.

Sottiaux told WIRED that when ChatGPT appeared in 2022, DeepMind had effectively been “sitting on” similar capabilities for almost two years. He said the launch made him want to move to San Francisco and find a way to work for OpenAI.

That comment is revealing. Sottiaux had already spent years around extremely advanced models. What impressed him was the decision to put one directly in front of millions of users.

His behavior at OpenAI has followed that same bias toward shipping. With Codex, he became unusually visible for an engineering leader, responding directly to developers, discussing product changes publicly and sometimes resetting usage limits when adoption milestones were reached.

The connection between DeepMind and OpenAI is pretty clear: Sottiaux seems much more interested in what happens after a model becomes good enough than in keeping the capability inside a research lab.

Q5What did Thibault Sottiaux work on when he first joined OpenAI?

When Thibault Sottiaux joined OpenAI, he initially built tools for OpenAI’s own researchers rather than jumping straight into ChatGPT.

WIRED reports that his first work resembled what he had already done at DeepMind: improving the infrastructure researchers used to do their jobs. That was a natural fit for someone who had spent years building ML workflows and model systems.

Coding agents quickly became relevant to that problem. If OpenAI’s models were becoming good enough to write and understand substantial amounts of code, the company could use them to speed up the people building the next models.

OpenAI already had several experiments moving in that direction. Some researchers were exploring agents that could work through codebases, while other teams were building systems able to execute commands and interact with computers.

Sottiaux eventually became one of the people who pulled those ideas into a product effort. His move into product leadership started from a very practical internal question: can an AI model actually do useful engineering work rather than simply suggest code?

Q6How did Thibault Sottiaux end up leading Codex?

Thibault Sottiaux became the leader most associated with modern Codex after OpenAI turned several coding-agent experiments into one serious product push.

The timing was important. Cursor was growing quickly, Anthropic was pushing Claude deeper into software development, and OpenAI already had strong coding models but a less coherent agent product around them.

Sottiaux helped bring together work that had been happening across different OpenAI teams. The resulting Codex could inspect repositories, modify files, run commands, test its own changes and keep working through a task.

That was a much bigger step than the original Codex model OpenAI introduced years earlier. The original product helped generate code. The new Codex was built around completing engineering jobs.

Sottiaux then became unusually close to the developer community around the product. He answered complaints and questions publicly, discussed usage limits and repeatedly interacted with heavy users. That direct feedback loop gave OpenAI something it often lacked with earlier agent experiments: a group of users who knew exactly what they wanted the agent to do and could quickly tell the team when it failed.

Q7Did Codex actually become a big product under Thibault Sottiaux?

Yes, Codex became one of OpenAI’s fastest-growing products while Thibault Sottiaux was leading it.

The early numbers already showed unusual momentum. OpenAI said Codex daily usage grew more than tenfold during one stretch after launch. It later reported more than three million weekly developers, then four million only two weeks later.

The product subsequently crossed five million weekly users. OpenAI also found that roughly 20% of Codex users were already knowledge workers rather than developers, with that group growing more than three times as fast as developers.

More recently, Sottiaux publicly said Codex had reached 20 million active users. He also referenced the 20 million milestone while discussing ChatGPT Work adoption with TechCrunch. OpenAI has used slightly different descriptions around that number publicly, so we would avoid treating it as a clean 20-million-Work-users metric. What is clear is the order of magnitude: the agent product family has gone from a developer experiment to tens of millions of users.

OpenAI’s own internal results strengthen the case. At one point, nearly all OpenAI engineers were using Codex, and the company reported that engineers were merging 70% more pull requests each week after adoption expanded.

Q8Is Codex still mainly a coding tool today?

Codex still has a large developer base, but its fastest expansion now comes from people using the same agent architecture for broader knowledge work.

OpenAI reported that knowledge workers had reached about 20% of Codex users and were growing more than three times faster than developers. Those users were creating reports, spreadsheets, presentations and contracts, while also using Codex for research and data analysis.

The company’s latest enterprise data makes the shift much harder to dismiss as a few unusual early adopters. Among enterprise customers, Codex recently generated 64% of the combined output tokens produced by Codex and ChatGPT.

The growth by department is even more striking. OpenAI says weekly active enterprise Codex users grew 108-fold in legal, 41-fold in sales, 41-fold in recruiting and 26-fold in marketing over the period it measured. Engineering grew fivefold.

Inside OpenAI itself, the average lawyer or recruiter now generates more than 85% of their AI output tokens through Codex. The average engineer is around 99%.

The interesting development now is no longer whether a coding agent can occasionally help a non-coder. In several groups, the agent interface is already becoming the main way people use AI at work.

Q9Why did OpenAI put Thibault Sottiaux in charge of ChatGPT too?

OpenAI expanded Thibault Sottiaux’s role because the company increasingly wants ChatGPT to behave like an agent, and Codex gave him a working model for how that can happen.

WIRED reported that Sottiaux was promoted to lead core products while OpenAI was combining ChatGPT and Codex into a broader product. OpenAI’s own public material now gives him the Head of Product & Platform title.

Look at what changed in ChatGPT. ChatGPT Work uses Codex technology to handle longer tasks, work across apps and files and create finished outputs such as documents, spreadsheets and presentations.

OpenAI had tried narrower agents before. Operator and the later ChatGPT Agent struggled to get broad adoption, and Sottiaux has said the technology was simply too early. Models needed more restrictions because they were less reliable.

Codex gave OpenAI a much stronger starting point. Developers had already taught the company how people behave when they genuinely trust an agent with real work.

Sottiaux now has to transfer that behavior from software engineering to almost everyone else.

Q10Is Thibault Sottiaux trying to turn ChatGPT into an agent for everything?

Yes, Thibault Sottiaux is currently helping turn ChatGPT into a personal agent that can handle far more than conversations.

OpenAI has repeatedly called the destination a “super app,” although that phrase can be misleading. The company is not simply trying to cram email, payments, coding and productivity tabs into one giant application.

Sottiaux told WIRED that the goal is to build the “world’s best personal agent” and make ChatGPT increasingly proactive. He has described a system that understands a person’s goals and preferences, connects to existing software and completes work through those services.

Codex provides much of the underlying architecture. A user may ask for a result in normal language while the agent quietly writes code, calls APIs, searches the web or manipulates files behind the scenes.

ChatGPT Work is already an early version of that idea. It can stay on a project for hours, break it into smaller tasks and create finished work. The important change is that the user increasingly describes the outcome while the agent decides how to get there.

Q11What does Thibault Sottiaux mean when he says “scaffolding is coping, not scaling”?

When Thibault Sottiaux says “scaffolding is coping, not scaling,” he means AI products should rely more on increasingly capable models and less on piles of custom logic built to compensate for weak models.

He discussed the idea in a Dev Interrupted podcast while explaining lessons from Codex. AI teams often surround models with complicated prompts, routers, fixed workflows and special rules to force reliable behavior.

Those tricks can work extremely well today. Sottiaux worries about building too much of the product around weaknesses that the next model generation may simply remove.

Codex offers a useful example. Instead of programming a separate rigid sequence for every possible coding job, the agent can increasingly inspect the situation, choose tools, run code, observe what happened and decide what to do next.

There are still plenty of hard engineering layers around that intelligence: permissions, sandboxes, tool execution, security, context and interfaces. Sottiaux’s argument concerns how much developers should micromanage the model’s reasoning itself.

It is a fairly radical product philosophy when applied to ChatGPT. Better models can suddenly make whole pieces of product logic unnecessary, which means OpenAI has an incentive to keep its interfaces unusually simple.

Q12Does Thibault Sottiaux want ChatGPT to make more decisions for users?

Yes, Thibault Sottiaux currently favors an AI experience where users state what they want and the system handles more of the intermediate decisions itself.

A recent TechCrunch interview made that philosophy unusually explicit. The interviewer cited Wharton professor Ethan Mollick’s comparison between ChatGPT Work, which tries to feel almost magical, and Anthropic’s Claude Cowork, which exposes more choices to the user.

Sottiaux embraced the simpler approach and pointed to adoption as evidence that people were responding to it. He described ChatGPT Work as powerful without requiring users to understand the machinery underneath it.

That connects to another idea he discusses frequently: discovery. OpenAI itself often discovers useful behaviors only after a stronger model ships. GPT-5.6, for example, became much better at handling large collections of documents, making slides, writing reports and doing research. The product team can then watch how people use those capabilities and build around what works.

This produces a very different product-development loop from normal software. OpenAI cannot completely specify the feature set first because each model release changes what the underlying system can realistically do.

Q13Is Thibault Sottiaux betting that AI agents will stop waiting for prompts?

Yes, OpenAI is currently testing the kind of persistent, proactive agent that fits directly with Thibault Sottiaux’s vision for ChatGPT.

The freshest evidence comes from code reviewed by WIRED. OpenAI has been testing a Codex “Persistent mode” that can continue working until the user puts it to sleep, rather than automatically ending after a task.

The experimental system goes further. Its instructions tell the agent to generate follow-up tasks for itself, work across sessions and use previous interactions and knowledge about the user when deciding what to do next. It can even message the user proactively, although the instructions tell it to do so sparingly.

OpenAI confirmed to WIRED that the feature is being tested and said there were no immediate plans for a broad launch. So Persistent mode is still an experiment rather than a finished ChatGPT feature.

The direction is hard to miss. Sottiaux previously described ChatGPT becoming “delightfully proactive,” and the new code is an early technical version of exactly that idea.

For now, the difficult problem is trust. A model making suggestions is easy to stop. An agent that keeps working, remembers context, runs tools and creates its own follow-up tasks needs much tighter control over what it is allowed to touch.

Q14Are normal workers actually using the AI agents Thibault Sottiaux is building?

Yes, AI-agent usage is now spreading well beyond software engineers, although developers remain the clearest proof that the model works.

OpenAI originally had more than five million weekly Codex users, with over one million already using it for work outside software development when ChatGPT Work launched. The broader Codex user base has since moved into a much higher range.

More useful than the headline user count is what people actually do with the product. OpenAI’s research found that more than 70% of Codex users had asked it to handle at least one task estimated to take a human more than an hour. Some heavy users now run many agents in parallel, effectively creating more agent working hours in a day than one person could physically perform.

The enterprise department data points in the same direction. Legal, recruiting, sales and marketing usage has lately grown much faster than engineering usage from a smaller base.

There is still a large gap between trying an agent and reorganizing a job around one. But we have enough evidence now to say that general-purpose agent usage has escaped the developer community.

Q15Does Thibault Sottiaux think AI agents will kill SaaS?

For now, Thibault Sottiaux seems to expect AI agents to change how people interact with software much faster than they eliminate the software itself.

At VivaTech, the discussion around Sottiaux and OpenAI engineer Peter Steinberger turned directly to the predicted “SaaS apocalypse.” Their answer was considerably more practical than the usual AI hype.

Reliable software still contains databases, permissions, business logic, compliance systems and years of operational work. Generating a disposable application from a prompt can be cheap; maintaining a dependable service is a different problem.

Sottiaux’s longer-term idea is more disruptive at the interface level. He described a personal agent that knows a user’s preferences and can generate custom software, dashboards and experiences on demand.

If that works, people may spend much less time clicking through the interfaces of individual SaaS products even while those products continue operating underneath. ChatGPT would become the layer through which the user reaches them.

That is a more plausible near-term threat to software companies: losing control of the customer interface before losing the underlying workload.

Q16What kind of work does Thibault Sottiaux think AI agents will take first?

Thibault Sottiaux expects AI agents to spread first through annoying, repetitive work that people already want to hand off.

At VivaTech, he used a simple personal example: he codes every day but dislikes writing tests, so that is exactly the kind of job he wants an agent to take.

He extended the same logic to ordinary office work. Rather than beginning with the most prestigious part of someone’s job, agents can handle the necessary tasks people keep postponing.

OpenAI’s current usage data fits that pattern surprisingly well. Knowledge workers use Codex heavily for data transformation, research, document production, workflow automation and retrieving information scattered across systems.

That also helps explain why adoption can creep up on people. Someone may first delegate a boring spreadsheet transformation, then research, then a recurring workflow, and eventually a project that would previously have required another person.

Sottiaux’s bet is essentially behavioral: trust grows through small completed jobs, and the size of the jobs increases afterwards.

Q17How much power does Thibault Sottiaux really have inside OpenAI?

Thibault Sottiaux has become one of OpenAI’s most influential product builders, but Greg Brockman and Sam Altman still sit above him in the company’s overall power structure.

Sottiaux recently confirmed that he reports to Brockman. Brockman has taken a much larger role across OpenAI’s products and operations, while Altman remains CEO and controls the broader corporate direction.

Sottiaux’s power is more specific. His organization touches ChatGPT, Codex, the API, enterprise products and agent infrastructure. Few OpenAI executives sit across that many parts of the company at once.

His influence also shows up in decisions beyond individual features. He has represented OpenAI publicly around acquisitions and integrations involving developer tooling and agent infrastructure, areas that shape what Codex and future agents can actually do.

The position becomes especially important as OpenAI tries to reduce the separation between its products. When ChatGPT, Codex, agents and the API share more technology, whoever controls that common product layer gains influence over a larger percentage of the company.

So Sottiaux is unlikely to decide OpenAI’s corporate future by himself. He can, however, have an enormous effect on what OpenAI’s technology feels like when people actually use it.

Q18Is Thibault Sottiaux basically an engineer or a product executive now?

Thibault Sottiaux is still an engineer by background and instinct, but his current job is unmistakably product leadership.

His résumé contains almost none of the traditional steps we would expect from a career product manager. Before OpenAI, he spent years as a software engineer and technical leader working on ML infrastructure, reinforcement learning systems, model serving and human data.

Even his public language remains highly technical. He talks about harnesses, model capability, inference costs, tool use and agent architecture far more often than about the usual product-management vocabulary.

At the same time, the problems he now owns are classic product problems at enormous scale: which capabilities users can understand, how much autonomy they will tolerate, what should cost $20 rather than hundreds of dollars, where ChatGPT ends and Codex begins, and how much complexity the interface should expose.

That combination is probably part of why his influence has risen so quickly. Frontier AI products increasingly change whenever the underlying models change. Understanding the model deeply is becoming unusually useful for the person deciding what the product should become.

Q19So who is Thibault Sottiaux, really?

Thibault Sottiaux is currently one of the people most directly responsible for turning OpenAI’s increasingly capable models into products that can actually do work for people.

His path explains why OpenAI gave him that responsibility. He spent years building AI infrastructure at Google DeepMind, led human-data work for Gemini, moved to OpenAI after becoming frustrated by the gap between research capabilities and real products, helped turn coding agents into Codex, and then watched Codex spread rapidly beyond software engineers.

OpenAI has now put him in charge of the core product layer connecting ChatGPT, Codex, agents, the API and enterprise usage.

The freshest developments make his mission even clearer. Codex has reached tens of millions of users. Non-developer agent usage is rising much faster than developer usage in several categories. ChatGPT Work brings Codex-style execution to ordinary knowledge work. OpenAI is currently experimenting with agents that can stay active, remember context and create their own follow-up work.

The hard part starts here. Software engineering gave OpenAI an unusually good environment for agents because code can be run, tested and checked. Everyday work involves more ambiguity, judgment, taste, permissions and real-world consequences.

Sottiaux’s importance now depends on whether the Codex approach survives that jump. If OpenAI succeeds in turning ChatGPT from something people ask questions into something they routinely trust with work, he will have been one of the main builders behind that transition.