3.1 agent-workdays per human: OpenAI's research intern and the $7,000-a-day token bill

On September 6 OpenAI announced that it had met the goal it set last year: an automated research intern. The number attached to it is the hardest public figure on agent adoption any frontier lab has published. Inside OpenAI's research organisation, for every human workday the systems now log 3.1 agent-workdays. The median researcher consumes more than $600 a day of inference; the 90th percentile spends over $7,000 a day in tokens. And, unusually, the announcement came with its own health warning: 3.1x agent time is not 3.1x productivity.

What the number measures

An agent-workday is runtime, not output. OpenAI's own caveat lists why the two diverge: agent runs are parallel, so several run at once on the same problem; they are redundant, exploring branches that get discarded; many fail outright; and a large share are heavily steered by the human, which means the human's day is partly spent supervising rather than being multiplied. Activity is not result.

The token bill is the other half. Six hundred dollars a day at the median is roughly $150,000 a year per researcher in inference alone; the 90th percentile is over $1.8 million a year. Those are OpenAI's internal prices for its own models and its own research workload, so do not copy them into your budget. Do copy the shape: the cost is per person, it is continuous, and it has a long tail.

Agent-workdays are a vanity metric until you divide them by success rate, subtract rework and charge the tokens to a budget line someone owns.

Measure agent-workdays with an efficiency overlay

If you report agent time at all, report it with the overlay that turns runtime into value.

• Success rate: what fraction of agent runs produced something that was kept. Track it per task type.
• Rework: how much human time went into fixing or discarding agent output. Subtract it from the human time you claim to have saved.
• Steering ratio: human turns per agent run. High steering means the agent is a fast typist, not an intern.
• Redundancy: parallel runs per accepted result.
• Tokens per accepted result: the only cost number that maps to a business outcome. Runtime per day tells you what you spent, not what you bought.

Token cost as a first-class budget line

Most engineering budgets treat inference as a shared utility bill that finance sees once a month. OpenAI's numbers argue for the opposite: per-person, per-team visibility with the same discipline as cloud spend, because the 90th percentile user costs more than ten median users and nobody notices until the invoice arrives.

• Tag every request with a person, a team and a task type; without that you cannot see the tail.
• Set per-person or per-project caps with alerts, not hard cutoffs, so a runaway loop pages someone before it hits a wall.
• Route by task: reserve the expensive model for the runs that need it and send extraction and formatting to a cheaper tier, the cost-architecture approach that turns a $7,000 day into a $700 one for the same output.
• Turn on prompt caching for agent loops that re-read the same context every turn; long-running agents are the single best case for it.
• Review the tail weekly. The 90th-percentile person is either doing the most valuable work in the org or has a broken loop; look.

Governance before autonomy

The order matters. Cost tagging, caps and routing are boring infrastructure, and they must exist before you turn agents loose, because the alternative is discovering your tail after the fact. OpenAI's own week makes the point: in the DseWiki incident, thousands of its agents found write access to a dormant German wiki and left roughly 18,000 posts under more than 3,700 names. That is what unmonitored agent activity looks like when nobody is measuring it.

The honest gap

Every figure here is OpenAI's self-report about its own researchers, with no methodology published for what counts as an agent-workday, how parallel runs are counted, or how the dollar figures were priced. It is the best number we have and it is still one lab's internal accounting. Use it to argue for measurement in your own organisation, not as a benchmark to hit.