OpenAI's researchers now run three agent-days for every human day
On 6th Sep’26, OpenAI published internal data on how its research organisation actually works. By mid-August the org was using 3.1 agent-workdays of effort for every workday of human labour. The median researcher now spends more than $600 a day on inference at API prices; the 90th percentile spends more than $7,000. Experiments per researcher hit an all-time high in August, but more than half of the successful four-to-eight-hour tasks still needed a person to step in at least once. High-level planning remains a minimal fraction of what the agents produce, and OpenAI itself says the overall pace of research will not keep up with these numbers.
What it actually means
This is the first honest budget for an agent-native team, published by the people furthest down that road. Two numbers to keep.
$600 a day per head is, over a working year, a salary in tokens, so "the agents do the work" does not mean the labour line disappears; it moves from payroll to the API bill. Also, more than half of the long tasks that succeeded needed a person to step in, while the planning stayed almost entirely human. If you are pricing an agent-heavy workflow, price the person who steers it as well as the tokens. The ratio is the headline. The intervention rate is the operating model.
The labour didn't disappear. It moved from payroll to the API bill.
$600/daywhat OpenAI's median researcher now spends on agents, at API prices
Source: OpenAI · Tuesday 8 September 2026