OpenAI Product Lead: ChatGPT Work Tops 20 Million Users

Thibault, OpenAI's head of core products, says ChatGPT Work has crossed 20 million users. The figure came out in an interview with TechCrunch — Thibault also runs the Codex coding tool and reports directly to Greg Brockman. The conversation covered why ChatGPT Work matters, how it's priced, and the underlying product philosophy of building agents that do work on people's behalf.

From Codex to ChatGPT Work: the next stage of rollout

Among Codex users, Thibault is known as the guy who suddenly lifts token limits — he regularly posts on X about usage resets and actually follows through with weekly resets. Codex originally launched for a fairly forgiving population of technical users, and Thibault says OpenAI is now working to carry that same capability over to non-developers. He framed the effort as reaching the widest possible audience with the technology, which is why ChatGPT Work is folded into the $20-a-month Plus plan rather than sold separately.

Minimal screen, maximum autonomy

The product philosophy Thibault stressed is staying out of the model's way rather than boxing it in. He said he'd rather resolve things through a single conversation than clutter the screen with buttons, pointing to ChatGPT Voice as an example — growth picked up noticeably once voice conversation was added to what had been a text-only product.

That direction sets OpenAI apart from its rivals. Wharton professor Ethan Mollick, in a Bluesky post, compared ChatGPT Work's aim for a "magical" experience against Anthropic's Claude Cowork, which puts choices in front of users at every step, almost like an A/B test. Asked about the comparison, Thibault said, "The world seems to be ready," pointing to the 20-million-user figure as evidence.

Pricing keeps moving

Thibault said OpenAI pushes for efficiency gains every day, and pricing has fallen 80% since Luna was introduced. He described this not as a one-off discount but as an ongoing recalibration, with the goal that six months from now, the same spend should buy noticeably more.

Item	Detail
User count	20 million (as of announcement)
Pricing	$20/month Plus plan
Price cut	80% since Luna
Core model	GPT-5.6
Ultrafast speedup	10x with fewer tool calls, 3-4x with more
Internal capacity policy	Ultrafast not given to all staff; external customers prioritized

Why Ultrafast isn't always 10x

The speed numbers vary depending on the task. Thibault said that for jobs with few tool calls but long output — writing lots of code for a website or a video game prototype, for instance — Ultrafast can finish 10 times faster. But once tool calls pile up, the bottleneck shifts to the network or the agent's routing path, and the perceived speedup drops to 3-4x.

He also revealed something about internal practice. When the interviewer asked whether OpenAI employees, with unlimited tokens, just max out their compute and leave Ultrafast running all the time, Thibault said no — Ultrafast isn't given to everyone on staff. A large chunk of that capacity is deliberately held back for external users and customers, because employees left unchecked could swallow all of production's GPUs. Internally, they watch usage and cap it at a level they consider reasonable.

The conversation also touched on how he personally works. Thibault said, "I have ADHD, so I'm always switching context," and that he draws energy from making lots of small decisions. That put him at odds with the interviewer, who said he prefers to focus on just two or three things at a time and welcomed Ultrafast for that reason. Still, Thibault admitted that when he does want to stay locked onto one thing, Ultrafast helps keep him from breaking flow.

GPT-5.6 and product as discovery

Thibault said every jump in model capability sends the team back to rediscover what's newly possible. With GPT-5.6, things like processing huge documents, generating high-quality slides and reports, and Deep Research-style tasks moved a notch closer to real work, and OpenAI builds product around wherever that bar lands. He called this part of iterative deployment — gathering community feedback and continuously refining from there.

Can you disrupt yourself once you have a cash cow

The interviewer asked whether it gets harder to keep a culture of self-disruption once a company matures and starts printing cash from an existing business. Thibault's answer centered on time: wherever AI is headed, and however humanity ends up benefiting from it, "it doesn't wait around for whatever you built up over the last month or three." So the key, he said, is to keep your eyes open, watch where things are moving, and position yourself to ride that wave.

What came next was the most candid, practical admission in the interview. "We ourselves only discover a model's capabilities after we've trained it," he said. "Benchmarks don't tell you everything." It takes real hands-on time with a model before you find out it can be used in some new way — and that discovery changes how the team thinks about the product itself.

The example he gave was the new voice feature. ChatGPT's voice has gotten much more natural and can now make tool calls, and Thibault said he personally now spends far more time talking to it than typing. Transcription quality in particular has improved enough that instead of typing prompts, he just leaves his phone out in the morning and talks.

Why open up email and iMessage too

The same day as the interview, OpenAI shipped a new feature letting ChatGPT access iMessage. Asked about concerns over this kind of expanding access, Thibault said the key is choosing a safe model. He argued that his company's models lead the field on safety, pointing to investment in the safety stack and transparent benchmark disclosure as evidence.

Editor's take

What really stands out in this interview is that Thibault didn't dodge the Claude Cowork comparison. Anthropic has chosen to build trust by showing users their options; OpenAI has gone the opposite way, minimizing intervention so the experience feels magical. Both companies are solving the same problem — getting non-developers to trust agents with real work — from opposite directions, and this looks like it's setting up a real test of which UX wins in the coming office-agent market.

There's also an interesting generational angle here. Codex first rolled out to a forgiving population of developers; now that same capability has migrated directly into ordinary office work like document writing and slide creation. Where automation once mainly benefited people who could write code, GPT-5.6 has effectively put the same kind of automation into the hands of anyone writing a report.

From a practical standpoint, two things are worth watching. One is pricing: if the ongoing price cuts Thibault described actually continue, teams currently worried about ChatGPT Work's token consumption could well be doing a lot more work for the same spend within six months. The other is data access — with email and iMessage integration now part of the picture, organizations would do well to set internal rules now for which accounts get connected to agents.

In the coming weeks, expect Claude Cowork to respond with its own adoption numbers in answer to OpenAI's 20-million milestone. The fact that Thibault directly engaged with the competitor comparison in this interview is itself a sign that OpenAI is well aware this is becoming a numbers fight.