OpenAI says Codex now generates 99.8% of its weekly internal output tokens

The numbers OpenAI is reporting on itself

The headline figure is about OpenAI, not its customers: the company says Codex now accounts for 99.8% of weekly output tokens generated inside OpenAI, and that the average engineer produces 99% of their output tokens through Codex rather than ChatGPT. Through August 2025, by contrast, the average OpenAI worker spent less than 10% of their tokens on Codex.

The paper stacks other internal measures on top of that: median combined output tokens by June 2026 were reportedly 56x their November 2025 level in Research, 32x in Customer Support, 27x in Engineering, and 13x in Legal. Non-developer weekly users grew 137x for individuals, 189x for organizations, and 12x inside OpenAI since August 2025.

Worth keeping in view: the task-horizon estimates come from an LLM-as-judge reading Codex transcripts, and the individual-user thresholds are drawn from a 0.1% random sample. OpenAI itself calls these figures directional rather than exact. So the strongest evidence here is behavioral — which tool people open — rather than a clean measurement of hours of human labor displaced.

The interface moves from a chat turn to a run you supervise

The framing OpenAI leads with is that agents change 'the unit of knowledge work from single interactions to delegated, long-horizon tasks.' That is a frontend claim as much as a model claim. A chatbot's surface is a turn: you type, you read, you type again. An agent's surface is a run that operates for minutes or hours while calling tools and iterating.

The report says that by May 2026, 80.6% of sampled individual users had made at least one request estimated at over 30 minutes of human work, 70.2% over an hour, and 25.6% over eight hours. If that's true, the interface a person spends their day in is no longer a message box. It's a queue of long-running jobs with intermediate states, partial results, and points where a human has to accept, reject, or redirect.

The most extreme data point makes this concrete: at the 99th percentile, OpenAI says users regularly generate more than 60 hours of Codex agent turns per day, spread across multiple parallel agents. Sixty agent-hours in a calendar day is only possible if the human is orchestrating, not conversing. The open design problem is what that person actually watches — because no one reads 60 hours of transcript.

Non-technical staff crossing into engineering work

The claim likely to travel furthest is that Legal, Finance, and Recruiting at OpenAI crossed into Codex being their primary AI tool around April 2026, later than Engineering but faster, and that the average lawyer or recruiter now generates more than 85% of their output tokens on Codex.

The occupation-versus-work heat map is more measured than the framing around it. For Product / Marketing / Ops, knowledge work is still the largest category at 51%, with engineering/coding at 25%. For Finance / Biz Ops, knowledge work is 34% and engineering/coding 31%. So non-developers are doing some coding-shaped work through Codex — automation, data transformation, debugging, structured analysis — but the data does not show them becoming engineers. It shows the cost of stepping into an adjacent, previously-blocked task falling far enough that they do it occasionally.

What a team building on this actually has to design

Read as a product signal from the vendor's own building, the announcement points at a specific gap. If the unit of work is a long, parallel, semi-autonomous run, then the hard part of an agent product is not the model call — it's the supervision surface. How does a recruiter or a lawyer, with no engineering background, review the output of an agent that just did an hour of technical execution they couldn't have done themselves?

That is the concrete implication for anyone shipping agent features: OpenAI's data describes a world where trust in an agent's result matters more than the agent's raw capability, because the person approving the work increasingly can't verify it line by line. The frontend problem OpenAI's own numbers imply — checkpoints, diffable results, cheap ways to reject and rerun — is the part they describe least, and the part a team copying this pattern will spend most of its time on.