OpenAI’s head of product says the company built its newest agent platform on purpose, not by accident: give a capability to technical users first, watch what they do with it, then repackage it for everyone else. That is the design philosophy Thibault Sottiaux described for ChatGPT Work in an interview published by TechCrunch on August 25, and it doubles as a roadmap for where OpenAI intends to take the rest of its product line.
Sottiaux, a member of technical staff who told TechCrunch he leads “all of core products,” meaning the API, agent infrastructure, enterprise tools, ChatGPT, and Codex, reports to OpenAI president Greg Brockman. He described ChatGPT Work, the agent platform aimed at white collar workers, as a deliberate export of Codex’s coding-agent capability to a non-technical audience, folded into the existing $20 a month Plus plan rather than sold as a separate product. “Codex was built for a forgiving technical audience,” he said, and OpenAI judged the underlying models mature enough to widen the audience “to as broad of a population as possible.”
He framed OpenAI’s product process as discovery-driven rather than planned from a fixed spec. “It is almost like a product of discovery as well,” Sottiaux said. “As we push on the frontier of capabilities of models, we also discover what it’s capable of, and we sort then lean into the things that it is the most capable of.” He cited GPT 5.6 as an example, saying the model’s step up in document processing, slide generation, report writing, and deep research shaped what ChatGPT Work became. That is a claim about how OpenAI decides what to build, not a benchmark result, and TechCrunch’s interview does not include independent verification of those capability gains.
Asked whether workers are actually ready for a product that tries to do everything at once, rather than the more incremental, choice-driven approach TechCrunch’s Tim Fernholz noted in rival Claude Cowork, Sottiaux pointed to adoption. “We definitely see that the world seems to be ready,” he said, citing 20 million ChatGPT Work users as OpenAI’s own reported figure. The interview offers no independent count to check that number against, and OpenAI’s history of citing raw user totals without engagement or retention data means the figure describes reach, not proof that the “magic” approach is winning.
The most useful admission came on cost. Pressed on the gap between what Plus subscribers pay and the token volume agent tasks consume, Sottiaux did not dispute that a gap exists. Instead he pointed to an 80 percent price cut tied to a model OpenAI calls Luna, calling it “a permanent price correction,” and promised that within six months users should get today’s capability “with less spend.” That is a forward promise, not a fix in hand, and it implicitly concedes that current Plus pricing does not yet cover what heavy agent use costs OpenAI to serve.
On the iMessage and email access ChatGPT Work now requests, Sottiaux offered a general safety assurance rather than specifics: “It’s important to pick models that are safe and aligned,” and OpenAI’s models are “world-class” on that front, he said, without naming a benchmark or third-party audit. Compare that to the specificity OpenAI has shown when it wants to claim a capability lead, and the gap in evidence stands out.
For operators evaluating ChatGPT Work against Claude Cowork or other agent platforms over the next quarter, the number to track is not the 20 million user figure but whether OpenAI’s promised token efficiency gains materialize on the Plus tier, since that determines whether the $20 price holds or migrates toward usage-based billing.
Based on an interview by Tim Fernholz published in TechCrunch on August 25, 2026.
