98% of OpenAI uses agents. 17% of its customers do. The gap is the actual product.

On July 9, OpenAI folded its Codex coding agent into the ChatGPT desktop app and launched ChatGPT Work, a $20-a-month tier that hands back finished documents, spreadsheets and small web apps instead of chat replies. The target is accountants, investors, doctors, and the ops person who builds the Monday metrics report by hand. OpenAI even built a benchmark for the ambition: GDPval scores model output across 44 occupations' worth of real knowledge work.

The growth is real. The Work and Codex apps went from roughly 10 million users in late July, per Bloomberg's reporting, to about 20 million by late August, and that same reporting of OpenAI's own numbers names people who do not code as the fastest-growing cohort. Set that against the billion-plus people who prompt ChatGPT and the shape of the bet is obvious: turn a chat habit into an agent habit, and charge for it.

So far, so launch-day. The interesting part is what OpenAI found when it pointed the instruments at itself.

In late June, OpenAI released a study with an academic byline that stretches across Columbia, Wharton and Duke: The Shift to Agentic AI: Evidence from Codex. Inside OpenAI, 97.9% of employees use Codex, up from about 40% in August 2025. Agents now generate 99.8% of the output tokens its staff produce across Codex and ChatGPT combined. The shift runs well past engineering: every department crossed over, and by April 2026 the majority of OpenAI's non-technical staff worked through Codex. The median employee in a legal role generated 13 times more monthly output tokens in June than in November. The median researcher, more than 50 times.

One honest caveat before you repeat any of this at a board meeting: every figure is self-reported by the company selling the product, and nobody outside OpenAI has audited them. Treat the internal numbers as a ceiling. The external ones are the part OpenAI had no incentive to publish, which is exactly why they are the part worth reading.

Every outlet that covered the study quoted the 17% as a curiosity. Read it as a gradient instead, because the three populations differ in exactly one way that matters: how much engineering sits around the model.

OpenAI's staff do not open a chat tab and hope. They work inside purpose-built harnesses, on tasks with checkable outcomes, under a company mandate, with a library of shared instructions. The study calls those instruction bundles skills, and the numbers split hard: 96.2% of internal Codex users invoke them, against 26.6% of organizational customers. More than 10% of users now run three or more agents at once each week, and the share of people submitting tasks an experienced human would need eight-plus hours to finish has grown nearly tenfold since January. Sophistication compounds where somebody built the scaffolding, and stalls where nobody did.

That reframes the whole product category. What vendors market as custom AI agent model development for non-developers mostly is not model development at all. The model is rented and identical for everyone. What separates 98 from 17 from 0.7 is whether anyone named the workflow, wired the context, and decided where the output lands. We keep seeing the same split in the survey data: 74% of leaders expect agents to redesign half their work while 21% are ready for it, and the production numbers across ten surveys cluster far below the adoption headlines. OpenAI just measured the same wall from the inside, with better instruments.

Here is the arithmetic OpenAI would prefer you not do at renewal time. TechCrunch's reporter ran ChatGPT Work casually for four days and burned about 80 million tokens, roughly $65 at list price, against a $20 monthly subscription. Four days of light use cost three months of revenue. OpenAI has already cut prices 80% for users of its Luna model to keep agents running. A price that loses money on an engaged user is a price built to create dependence, and prices built to create dependence move once the dependence exists.

None of that makes the seat a bad buy. It makes it a mispriced one, in your favor, for now. The planning mistake is budgeting agent-heavy workflows at seat price. Price the workflow at token cost, because that is the number the subsidy is currently hiding, and it is the number your finance lead will meet in eighteen months. We benchmarked what companies actually spend on AI agents across about 50 primary sources, and the run-cost line is the one buyers most consistently forget.

Take the pitch at face value for a moment, because the upside is real. A two-person ops team names its ugliest recurring deliverable, the weekly metrics report that eats a Thursday afternoon, and hands it to a Work agent with the source spreadsheets attached. The agent drafts it, a human checks it, and Thursday afternoon comes back. Multiply by the reconciliation, the dashboard upkeep, the research memo, and a $20 seat quietly absorbs the first hire you were about to make. The study's internal numbers show what the far end of that road looks like: legal staff at 13x their previous output, agents running while their owners are in meetings. And right now the road is subsidized. The cheapest agent capacity you will ever buy is being sold at a loss to win you.

That is the honest case for buying seats this quarter, and it costs us nothing to make: nobody pays us a cent when a team subscribes to ChatGPT. And if the blocker is that nobody can name the first workflow to hand over, that list is precisely what a 10x audit produces.

The failure mode is just as well measured, and it is the majority outcome. Week one, someone demos an agent building a dashboard and the Slack channel lights up. Week three, the seats are quiet, because nobody named a workflow, nobody owns the output, and the first confident-but-wrong draft that reached a customer burned the team's trust. You are now in the 82.7% majority of organizational subscribers that does not touch the agent, still paying, and the renewal auto-bills while the price waits to normalize.

OpenAI's own gradient says the difference between the two outcomes is not the model, which is identical in both, and not enthusiasm, which both teams had in week one. It is whether anyone did the unglamorous work around the model: picked the workflow, wired the context it needs, decided where output lands, and routed exceptions to a person. Inside OpenAI, hundreds of engineers built that scaffolding for their colleagues. Your team gets a chat window. That scaffolding is the actual product. It is what we build when a team wants the 98% experience without employing the engineers, and what an engineer embedded in your team keeps running when the workflow will not sit still long enough to spec. It is also honestly not always worth paying for: if you can name the workflow, attach the context and add a review gate yourself, do that first and keep the money.