Thibault Sottiaux, who leads OpenAI’s core products including the API, agent infrastructure, ChatGPT and Codex, told TechCrunch that ChatGPT Work has reached 20 million users since launch. He described the product’s design philosophy as deliberately minimal: get out of the way of the model rather than build interfaces around it.
That is a specific bet against a rival approach. Sottiaux, asked about Wharton professor Ethan Mollick’s observation that ChatGPT Work “tries to be magic” while Anthropic’s Claude Cowork surfaces A/B style choices for users to make, did not dispute the framing. “We definitely see that the world seems to be ready,” he said, pointing to the adoption numbers as evidence that workers will accept an agent that acts autonomously rather than one that asks for input at each step.
The distinction matters because it sets up a live experiment in enterprise AI adoption. If ChatGPT Work’s low-friction, low-visibility design produces fewer errors and more renewals than Claude Cowork’s choice-heavy interface, OpenAI’s bet on minimal surfaces becomes the template other vendors copy. If it produces more silent failures in professional workflows, the choice-driven model gains ground instead. Neither company has published retention or error-rate data that would settle it.
Sottiaux, who reports to OpenAI president Greg Brockman, framed the product’s evolution as reactive rather than planned. “It is almost like a product of discovery as well,” he said, describing how the team leans into whatever capabilities a new model unlocks rather than designing features in advance. He cited GPT-5.6 as a step change for document processing, slide generation and report writing, the kind of work ChatGPT Work packages for non-technical staff.
On cost, Sottiaux pointed to a price cut tied to a model he called Luna, which he said cut prices 80 percent as a permanent correction rather than a promotion. He predicted that six months from now, users should get the same capability for less spend, extending a trend OpenAI has repeated with each new model generation. The claim is consistent with the company’s public pricing history, though OpenAI did not provide usage data tying the cut to specific ChatGPT Work workloads.
The interview also touched on OpenAI’s decision to extend ChatGPT into iMessage access, rolled out the same day as the conversation. Sottiaux’s answer leaned on OpenAI’s safety investments rather than specifics: “It’s important to pick models that are safe and aligned.” He did not describe what data ChatGPT retains from message access or how long it is stored, a gap that will matter more as agents move from drafting documents to reading personal communications.
The economic logic underneath all of this is that OpenAI wants to own the application layer, not just the model layer, before rivals do. Sottiaux confirmed as much when asked directly: the more utility ChatGPT provides, he said, the more willingly users pay for it. Bundling agent access into a $20 Plus subscription is a distribution strategy as much as a product one, and it puts pressure on Anthropic, Google and any enterprise software vendor selling a narrower, choice-driven agent at a comparable price.
For teams evaluating AI agent platforms this quarter, the practical takeaway is to test both interaction models directly: run the same workflow through a minimal-interface agent and a choice-driven one, and measure error rates and time saved rather than trusting adoption figures alone.
TechCrunch published this interview with OpenAI’s Thibault Sottiaux, conducted by Tim Fernholz, on August 25, 2026.
