ChatGPT Work: 98% of OpenAI staff use it, under 1% of outside users do

What ChatGPT Work actually is

ChatGPT Work launched last month as part of OpenAI's lowest-priced paid tier. The goal is to bring agents to white-collar workers who spend their days at a computer — accountants, investors, doctors. It's essentially a version of Codex reworked so that people who don't code can tap into the same kind of autonomous task execution developers already get from coding tools. We covered this on August 11 when OpenAI added a feature that automatically syncs work between ChatGPT Work and Codex, letting projects, conversations, skills, and plugins move between the two products — groundwork that was already being laid around that time.

In its launch post, OpenAI described its goal as "a world where AI helps everyone bring their biggest ideas to life, not just answer questions." That shift is already well underway in coding, but by OpenAI's own read, it's been slower to spread into other parts of the business.

Why OpenAI is pushing so hard

Longer-running agents burn through more tokens, which means better per-user revenue for OpenAI. Coding has been the most profitable corner of the AI industry so far, but justifying the massive spend on training and compute means expanding into professional work more broadly. Meanwhile, vertical-specific rivals like Harvey (legal) and Clay (sales) are going after the same customers by staying model-agnostic and swapping in whichever AI works best for the job. As a16z's Christian Catalini put it in a blog post, "if the large model companies don't quickly lock down the complementary assets needed to scale AI in the market, the value migrates elsewhere."

98% inside, 1% outside

According to an OpenAI-backed survey, 98% of OpenAI employees were using Codex as of June. Organizational subscribers came in at 17%, and individual subscribers at under 1%. That gap is both the problem OpenAI needs to solve and the opportunity it's chasing.

Group	Codex usage (as of June)
OpenAI employees	98%
Organizational subscribers	17%
Individual subscribers	Under 1%

Ambrosino recalled that when non-developers first started using Codex, it "was hostile — asking what code even is, telling you there's an empty diff." Since February, the team has been reworking it to be usable by anyone. OpenAI says it uses GDPval, an internal benchmark covering hundreds of knowledge-work tasks across 44 professions, combined with user feedback, to decide what to fix.

Putting it to the test

TechCrunch's own hands-on testing produced mixed results. Pulling a child's oddly formatted preschool schedule out of an email and moving it into Google Calendar worked well and saved a lot of manual entry. Tasks that used to require writing a Python script — like building a financial analysis dashboard for public companies or a space-launch database — got handled automatically too. But trying to grant "read-only" access to a cloud drive kept failing, and the system eventually said it would only work if given full access through the mobile app. What's supported and what isn't also felt inconsistent: you can create events on Google Calendar, but you can't create a new calendar altogether.

Feature	Current status
Create Google Calendar events	Supported
Create a new calendar	Not supported
Grant read-only permissions	Not supported (full access only)
Change key settings	Web app only

It's notable that Sam Altman himself reportedly uses the tool to plan his own vacations, which fits with the broader push to move beyond coding automation and into everyday task coordination. Akshay Nathan, who leads OpenAI's product engineering team, said that in an environment flooded with information, "the ability to look at something and act on it is actually the scarce part" — and that ChatGPT's value lies in making access people already have "actually usable."

The competitive landscape

ChatGPT Work isn't the only product chasing this space. Claude Cowork and Perplexity's browsing agents take a similar approach, stitching together email, browsers, and SaaS platforms to tap into a user's existing work context. Yet when asked how Codex differs from Claude Cowork, Joe Gershenson, who leads harness engineering at OpenAI, said plainly: "Honestly, I don't pay much attention to the harnesses other companies build." That's notable, since harness design — deciding what an agent can see and which tools it can reach behind the scenes — is shaping up to be the real battleground here, much like how CLI tools for coding reshaped how developers work in the first place.

Editor's take

The real bottleneck here isn't technical — it's evaluability. Code either runs or it doesn't. A good report or a good sales strategy doesn't resolve that cleanly. The moment OpenAI steps outside coding, it runs into a world where "what counts as correct" is hard to define, and the fact that it's leaning on benchmarks like GDPval to bridge that gap says a lot about how badly it wants a foothold in white-collar work.

Anyone who's tried wiring a coding agent into real workflows will recognize the permissions headaches described here — wanting read-only but only getting full access, settings that only change on the web. That kind of friction always shows up when a tool built for power users starts moving toward a general audience. It's the same pattern coding tools went through: a CLI was enough for early adopters, but reaching everyone else eventually required buttons and clear instructions.

For teams here evaluating whether to adopt agents like this, the 98%-versus-1% gap is worth reading not as a sign of failed adoption, but as a signal about onboarding design. Internal employees already had workflows that Codex could simply mimic; for outside users, getting stuck on permission settings became the first impression. Documenting the scope of access needed before rollout — and separating tasks that genuinely require read-only access from those that need full access — actually saves real time.

In the months ahead, harness design itself is likely to become a key axis of competition. As the gap in raw model performance narrows, what will really separate products is which screens and buttons let people actually put that performance to work.