TechCrunch visited OpenAI and came back with a set of numbers that don't add up the way you'd expect. As of this June, 98% of OpenAI employees had used Codex internally. Outside the building, only 17% of organizational subscribers were using agent-style tools, and among individual subscribers that figure drops below 1%. For scale, ChatGPT's prompt-based user base has already passed 1 billion, and its desktop app has been installed 20 million times.
A company betting its next product generation on agents has almost every employee using one — and almost no customers doing the same. That gap says more about where agents are actually stuck than any benchmark could.
Outside of coding, there's no way to grade the work
Engineers have a built-in way to judge an agent: does the code run, do the tests pass, does the pull request get merged. Whether the task succeeded is binary, and whether a user comes back a second time depends on whether the first run genuinely saved them time. Office work has no equivalent scoring system — whether a client brief reads well or a calendar got sorted correctly can only be judged by a human looking it over again, and that second look eats into whatever time the agent was supposed to save.
"Everything is coding agent shaped...the reason is that they only have training data for coding agent tasks."
Every agent ends up shaped like a coding agent, because the only training data vendors actually have is for coding tasks.
That's from Pi's creator Mario Zechner, and it flips the usual explanation on its head. Vendors haven't chosen developers out of preference — dense, labeled supervision signal simply exists only in code. GitHub alone holds billions of public "change → test → merge" records, each one carrying its own pass/fail label for free. Office work has no equivalent corpus, and nobody is out there labeling it. Training an agent to independently produce a usable meeting summary would first require millions of good and bad summaries laid out side by side — and right now, no one is doing that.
The permissions bar is higher than it looks
Andrew Ambrosino, a staff engineer on the desktop app, raised a specific worry: ask an agent to draft a document, and it might casually surface something from a private chat thread it happened to see along the way. Getting an agent to do real work means handing over read-and-write access to email, calendar, files, and chat tools. Getting people comfortable handing that over means the product first has to spell out exactly what the agent can and can't see. That layer of design costs far more than bolting on another tool integration.
OpenAI itself isn't aligned on this internally. Joe Gershenson, the engineering lead for harnesses, put it bluntly: "The honest answer is that I really don't look at the harnesses that they're building." Thibault Sottiaux, who leads core product, framed it closer to a mission statement: "It's the very mission of OpenAI — to bring everyone along."
Subsidies won't stretch to a billion people
There's also a math problem that hasn't been solved. TechCrunch cited one example: a user burned through 80 million tokens in four days, which pencils out to roughly $65 at list price — while that user was paying a $20-a-month subscription. Developers can live with that kind of inversion, since their willingness to pay is high, retention is strong, and the ceiling is well understood. Spread the same inversion across a billion ordinary users, and it becomes a gap that widens on its own.
OpenAI's response has been to push unit costs down — cutting prices 80% for users on the Luna model, for instance. Every price drop tends to be followed by faster growth in per-user usage, so the math ultimately has to be closed by model efficiency catching up, not by rate-limiting. Today's five-hour usage window on ChatGPT Plus is essentially a rule standing in until efficiency catches up. Enterprise customers get a more direct fix — seat-based tiers, separate pricing for heavy users, costs spelled out in the contract. Individual subscribers don't have that room to maneuver; $20 a month is the psychological price point, and moving it up a tier starts losing subscribers.
Ahead, but only by half a step
The competitive picture has flipped twice this year. Claude Code held a clear lead through April 2026, after which Codex clawed back a slight edge. On coding tasks specifically, the gap between the two is now small enough that only individual tasks can tell them apart — and the next round clearly won't be decided there. Whoever gets the 99% who don't write code to actually use an agent first is the one who turns this market from a few million developers into a billion people.
The distance between 98% and 1% isn't a model-capability problem. Closing it means getting someone who has never heard the term "permission scope" willing to hand their inbox over to a program that runs on its own — and not check its work line by line afterward. Authorization, execution, and skipping the check: none of the three is solid yet.
Sources: TechCrunch reporting based on interviews with the OpenAI team, CocoLoop, OpenAI's public pricing page; internal/external usage shares, desktop install figures, and token-cost figures are cross-checked against the original reporting and public pricing, with cost conversions flagged as rough estimates.