OpenAI report: median researcher's agents burn $2,400 of tokens a day

OpenAI report: median researcher's agents burn $2,400 of tokens a day

In a second report on internal research acceleration, OpenAI has published its own data on how deeply AI agents are embedded in developing new models. The company says it aims to document indicators of so-called recursive self-improvement (RSI), where AI systems speed up the development of future generations. The report says usage of internal tools like Codex and Dots has surged recently.

The usage numbers are stated at regular API list prices, so they are theoretical, not actual spending. On that basis the median researcher would have consumed $2,400 worth of tokens per day in October, up 62 percent from the previous month. Among the top 10 percent of users, that theoretical value doubled to $23,000 per day. Converted into eight-hour workdays, the agents reportedly run the equivalent of 5.7 workdays per researcher per calendar day, and about eight in the pre- and post-training teams. More than two-thirds of tracked researchers ran at least four agents in parallel at peak usage. In September, the theoretical API equivalent of that agent runtime was twice as high as researchers' regular salary costs, according to OpenAI. That calculation excludes equity packages, which make up 70 to 80 percent of total compensation at OpenAI. Measured against the company's total research compute, internal agent use accounted for just 2 percent.

Success rates are rising. Between March and October, the estimated success rate of autonomous agents working without human intervention on two- to four-hour tasks rose from 23 to 46 percent. For four- to eight-hour tasks, it climbed from 13 to 35 percent. A breakdown using an Epoch AI taxonomy shows the systems mainly handle simple tasks. Agents reportedly complete training-run monitoring 79 percent of the time and answer technical questions independently 72 percent of the time, but they frequently fail at more complex engineering work: the success rate is 29 percent for building evaluation and reinforcement-learning environments and 25 percent for writing code for inference and serving systems.

On how much staff value the tools: in an internal, non-representative survey of 101 employees, researchers said at the median that they would rather give up 40 percent of their team's staff or 30 percent of their compute capacity than go three months without current coding agents. In a separate quick poll of 28 employees, 27 preferred the latest internal model over a mid-career human engineer.

Qualitative interviews with ten researchers revealed the downside of heavy usage. The models lack any "taste" for good research ideas and for code and data quality, according to the interviewees. Agents get lost in unnecessary steps, make wrong assumptions in analyses, and submit subpar "AI slop" code changes along with unusable data artifacts. Close human oversight is essential to avoid wasted compute cycles. The key conceptual ideas for GPT-6 Astra and GPT-6.2 Astra came exclusively from humans in the post-training and safety areas, according to the teams interviewed. Only in pretraining for GPT-6.2 Astra did individual researchers say they gained important insights through dialogue with Codex.

OpenAI cautions in its conclusion that the data gathered so far from human-AI collaboration may not extrapolate linearly. It points to recent, rapid AI breakthroughs in mathematics and considers it possible that similarly abrupt advances could occur in deep-learning research itself as models gain more autonomy.

The Decoder, which reports on the document, adds context. The evaluation organization METR had recently estimated, after an eleven-week look inside competitor Anthropic, that AI use there had nearly doubled the pace of development. Before that, OpenAI had declared in its first research-acceleration report in early September that it had reached its self-imposed goal of an automated AI research intern. Chief scientist Jakub Pachocki warned of losing control in an accompanying post, after an audit revealed numerous gaps in the oversight system.

Key facts

  • At API list prices, the median OpenAI researcher would have consumed $2,400 of tokens per day in October, up 62 percent on the previous month; the top 10 percent of users reached a theoretical $23,000 per day.
  • Agents reportedly run the equivalent of 5.7 eight-hour workdays per researcher per calendar day, and more than two-thirds of tracked researchers ran at least four agents in parallel at peak.
  • Estimated autonomous success rates between March and October rose from 23 to 46 percent on two- to four-hour tasks and from 13 to 35 percent on four- to eight-hour tasks, but sit at only 29 percent for evaluation and RL environments and 25 percent for inference and serving code.
  • In a non-representative survey of 101 employees, the median researcher would give up 40 percent of team staff or 30 percent of compute rather than go three months without coding agents; 27 of 28 polled preferred the latest internal model over a mid-career human engineer.
  • Interviewees with ten researchers described a lack of research taste and 'AI slop' output; the key ideas for GPT-6 Astra and GPT-6.2 Astra came exclusively from humans in post-training and safety.

Why it matters

This is OpenAI's own first-party data on how much of its model development now runs through coding agents, published as part of its effort to document indicators of recursive self-improvement. It is the second such report; the first, in early September, declared the company had reached its goal of an automated AI research intern. The new numbers show fast growth (median theoretical token use up 62 percent in a month) alongside a clear limit: the agents handle routine work well but fail often on complex engineering, and the key conceptual ideas for GPT-6 Astra and GPT-6.2 Astra still came from humans. METR's recent estimate that AI use at Anthropic had nearly doubled its pace of development gives a comparable outside reading for a competitor.

Who it affects

Most directly, OpenAI's researchers and engineers, especially the pre- and post-training teams, where the agents reportedly run about eight workdays per researcher per calendar day. It also concerns anyone tracking how fast frontier labs can speed up their own development, and teams elsewhere deciding how much to lean on coding agents for research work.

How to use it

Treat the report as a benchmark for what heavy agent use looks like inside a frontier lab rather than as a recipe. The practical lessons from the interviews are that agents suit routine tasks such as training-run monitoring (79 percent) and answering technical questions (72 percent), that close human oversight is needed to avoid wasted compute, and that running several agents in parallel is common: more than two-thirds of tracked researchers ran at least four at peak. The source gives no pricing or access details for Codex or Dots.

How solid is it

These are self-reported figures from OpenAI, relayed by The Decoder. The token costs are theoretical API list-price equivalents, not actual spending, and the salary comparison excludes equity, which makes up 70 to 80 percent of total compensation. The 101-employee survey is described as internal and non-representative, the 28-person poll was a quick poll, and the interviews cover only ten researchers. The 5.7 workdays figure is described as 'reportedly', and the source does not say whether it is a mean or median. The success rates are described as estimates, and the source does not say how they were measured beyond that.

Risks and caveats

OpenAI itself cautions that the data from human-AI collaboration so far may not extrapolate linearly, and considers it possible that abrupt advances could come in deep-learning research as models gain more autonomy. Interviewees reported agents getting lost in unnecessary steps, making wrong assumptions in analyses and submitting low-quality code and unusable data. Internal agents used just 2 percent of total research compute, so the current footprint is small relative to the whole. The Decoder also recalls that Chief scientist Jakub Pachocki warned of losing control in a post accompanying the first report, after an audit revealed numerous gaps in the oversight system. The source does not say what those gaps were or who conducted the audit.

“Close human oversight is essential to avoid wasted compute cycles.”

— The Decoder, summarising interviews with ten OpenAI researchers