OpenAI says its median researcher now spends over $600 a day on inference

OpenAI published a set of internal measurements on 6 September 2026 describing how coding agents have changed the daily work of its own research organisation. The post gives figures for spend, concurrency, and how often agents finish what they are asked to do.

The spend numbers

OpenAI says that at the start of 2026 the median researcher, ranked by agent usage, was using coding agents only in modest amounts. By mid-August the median researcher was using more than $600 per day of inference at API prices. The company puts the 90th percentile user in the research organisation at more than $7,000 of tokens per day.

The company also gives a ratio for total effort. Before June 2026, it says, total agent runtime across the research organisation was still below total human labour. Measured in standard eight-hour workdays, as of mid-August the organisation used 3.1 agent-workdays for every workday of human labour.

Success goes up, supervision does not go away

OpenAI classified sessions by how long a human would have taken on the same task. It says success rates generally rose from January to July across several difficulty buckets. It states the limit in the same breath: over the last six months, more than half of successful tasks in the four to eight hour bucket involved one or more human interventions.

The company describes its own measurement as preliminary. It notes that its available compute also grew significantly since 2025, so the rise in experiments per experimenter is correlated with Codex adoption rather than shown to be caused by it. These are OpenAI's figures about OpenAI, with no independent audit behind them.

The quietest finding is about support channels

One observation has nothing to do with writing code. OpenAI says several internal teams that used to run office hours, where researchers came to get help debugging their experiments, have seen attendance fall through 2026, and one team stopped holding the sessions entirely. It plots daily posts to a main internal technical-support channel and reports a decline, adding that to its knowledge the traffic has not moved to another human-run channel.

That is a different claim from "agents write more code". It says agents absorbed the question a person used to ask a colleague.

The $600 figure is the one worth sitting with, because it is a per-person daily rate at API prices inside a company that has every reason to use its own tools heavily. Read as a ceiling rather than a target, it says that agent spend behaves like infrastructure cost and not like a tool licence. A team that plans for a fixed seat price per developer per month has budgeted for the wrong shape.

The intervention figure sets the other boundary. If more than half of the successful multi-hour tasks needed a human to step in, then delegating longer work does not remove the person, it changes what they do with their time. The reviewing, steering and judging is the part that stays, and it is the part that does not get cheaper as token prices fall.

Source

Research acceleration: The view inside OpenAI, OpenAI, 6 September 2026. Every figure in this item is OpenAI's own measurement of its own research organisation, as stated in that post.