OpenAI report shows median researcher's daily coding-agent spend rising to more than $600
A rare inside look at agent adoption shows spend, agent-days and human interventions all climbing together, and what each figure actually measures.
OpenAI has published a report titled "Research acceleration: The view inside OpenAI" describing how its own research staff are using coding agents, alongside a companion essay by Chief Scientist Jakub Pachocki called "An Alien Mind". Both were highlighted on 6 September by Simon Willison, who noted that the report frames the work around what OpenAI calls recursive self-improvement, or RSI, and does not expand the acronym.
The centrepiece is a chart of daily inference spending per researcher across 2026. As described by Willison, the line sits close to zero in February, rises to around $50 a day by April and roughly $150 by June, plateaus at about $150 to $165 into July, then climbs steeply to around $600 by late August. The report itself states that by mid-August the median researcher was using coding agents every day at more than $600 a day in inference, while a user at the 90th percentile was consuming more than $7,000 of tokens a day.
The report also expresses agent work in units of human time. As of mid-August, OpenAI says it was recording 3.1 agent-workdays of effort for every human workday, whereas before June total agent runtime had been below the equivalent human labour. Experiments per active experimenter reached an all-time high in August, measured against tracking that began in January 2025, and success rates on researcher tasks rose between January and July. More than half of successful tasks lasting four to eight hours still involved at least one human intervention.
Using a task taxonomy from Epoch AI, the report says all six categories of research work grew between January and August, with technical help and monitoring of training runs among the largest gains, while high-level planning stayed minimal. Willison speculates that the late-July inflection coincides with internal access to a newer model, but the report does not state that.
Why it matters
This is one of the few public datasets on how a frontier lab actually uses agents day to day, and it shows spending moving from tens of dollars to hundreds per person in a matter of months. The human-intervention figure is the counterweight: even inside the lab that builds the models, long tasks are still collaborative rather than fully autonomous. The Epoch categories suggest agents are taking on the plumbing and monitoring work first, leaving the planning to people. Anyone budgeting for agentic tooling should expect a similar curve, and a similar ceiling.
This is one of the few public datasets on how a frontier lab actually uses agents day to day, and it shows spending moving from tens of dollars to hundreds per person in a matter of months.
What you can learn from this
Inference spend is a direct proxy for how much an agent is doing. Language model APIs bill by token, and an agent that reads files, runs tests and rewrites code consumes tokens on every loop. A daily figure such as $600 therefore tracks the number and length of agent runs, not the price of a subscription. That is why the chart is a useful measure of adoption: it rises only when people hand the agent longer and more frequent jobs. Comparing the median with the 90th percentile also shows how uneven adoption is, since a few heavy users can dominate the total.
Agent-workdays are a way of comparing machine and human effort. If an agent runs continuously for 24 hours, that is roughly three eight-hour shifts, so a ratio of 3.1 agent-workdays per human workday means agents are running for longer than the people supervising them. Two agents running in parallel for a day would count as roughly six. The unit works because agents do not stop at the end of the day, but it says nothing about quality. It is best read alongside success rates and intervention counts.
Long-horizon tasks fail differently from short ones. A four-to-eight-hour job involves many decisions in sequence, and a small error early on compounds through everything that follows. That is why more than half of such successful runs still needed a human to step in, even at a lab with unusual model access. Checkpointing progress, reviewing intermediate results and keeping tasks decomposable are the standard ways to keep that compounding in check. The intervention count is therefore a design signal, not a failure grade.
Task taxonomies make adoption measurable. Splitting research work into categories such as technical help, monitoring and planning lets an organisation see which parts are being automated first. The pattern reported here, with plumbing and monitoring growing fastest and planning barely moving, matches how agents behave: they excel at well-specified, verifiable work and struggle when the goal itself is ambiguous. Measuring by category tells you where to invest and where to keep humans in the loop, and it gives you a baseline to compare against next quarter.
Recursive self-improvement is a research programme, not a switch. The idea is that models help build better models, which then help more, but in practice that loop runs through humans who decide which experiments to run and judge the results. The data here shows exactly that: more experiments per person and more agent time, with people still directing. Understanding it as accelerated human research, rather than autonomous research, keeps expectations grounded.
We teach this
How to use this in practice
Instrument your own agent spend for one week. Pick whichever coding agent you use, export or record its token usage each day, and convert to dollars using the published price for the model. Done is a seven-point chart, however rough, with a note on which day was highest, what you were doing that day, and the model price you used. If your tool does not expose usage, estimate from the number of runs and their average length, and label the estimate as such. Compare the shape of your line with the report's: flat, then a step up when you started giving the agent longer jobs.
Compute your personal agent-workday ratio. Log how many hours agent runs were actually executing versus how many hours you worked, for three days. Keep the log simple; a text file with start and stop times is enough. Divide the two. Done is a single number and one sentence about whether you were the bottleneck or the agent was. Then note how many of those runs finished without you stepping in, which is your equivalent of the report's intervention figure for long tasks.
Classify one day of tasks with a simple taxonomy. Write down every job you gave an agent and tag each as technical help, monitoring, experiment setup, analysis, writing or planning. Count which categories succeeded without your intervention. Done is a small table that shows where you can hand over more and where you should not. Planning tasks will probably be the smallest group, matching the pattern OpenAI reports.
Draw the loop as a diagram. Sketch four boxes in a ring: human sets task, agent executes, human reviews and intervenes, results feed the next task. Annotate each arrow with the figure from the report that applies. Add a fifth box for cost, fed by the execute step, so the diagram also shows why the spend line and the agent-days line move together. Done is a picture that explains why spend, agent-days and intervention rates all belong on the same page.
Sources
- Research acceleration: The view inside OpenAI — Simon Willison
- Research acceleration: The view inside OpenAI — OpenAI
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report