OpenAI report: median researcher's agent usage would cost $2,400 per day at API list prices in October
OpenAI's second internal research acceleration report says the median researcher would have consumed $2,400 of tokens per day at API list prices in October, up 62 percent from September. Agent success rates on longer tasks have risen, but the systems still fail at complex engineering work and need close human oversight.
OpenAI says the median researcher at the company would have consumed $2,400 worth of tokens per day in October at regular API list prices, up 62 percent from the previous month. The figure comes from the company's second report on internal research acceleration, covered by The Decoder on Nov 19, 2026. OpenAI says it aims to document indicators of recursive self-improvement (RSI), where AI systems speed up the development of future generations.
Usage figures
According to the report, usage of internal tools such as Codex and Dots has surged recently. Among the top 10 percent of users, the theoretical value doubled to $23,000 per day.
- Converted into eight-hour workdays, agents 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 agent runtime was twice as high as researchers' regular salary costs, according to OpenAI. That comparison excludes equity, which makes up 70 to 80 percent of total compensation at OpenAI.
- Internal agent use accounted for 2 percent of the company's total research compute.
Success rates
Between March and October, OpenAI's estimated success rate for autonomous agents working without human intervention rose from 23 to 46 percent on two- to four-hour tasks. For four- to eight-hour tasks, it rose from 13 to 35 percent.
A breakdown using an Epoch AI taxonomy shows the systems mainly handle simple work:
- Training-run monitoring: 79 percent success
- Answering technical questions independently: 72 percent
- Building evaluation and reinforcement-learning environments: 29 percent
- Writing code for inference and serving systems: 25 percent
Employee surveys
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.
Limits reported by researchers
Qualitative interviews with ten researchers pointed to a lack of "taste" for good research ideas and for code and data quality. Interviewees said agents get lost in unnecessary steps, make wrong assumptions in analyses, and submit subpar "AI slop" code changes and unusable data artifacts. Close human oversight is needed to avoid wasted compute.
Teams interviewed said the key conceptual ideas for GPT-6 Astra and GPT-6.2 Astra came exclusively from humans in the post-training and safety areas. Only in pretraining for GPT-6.2 Astra did individual researchers say they gained important insights through dialogue with Codex.
OpenAI cautions that 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 as models gain more autonomy.
The report follows OpenAI's first research-acceleration report in early September, in which the company declared 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.
What this means
The data shows heavy, growing dependence on agents inside OpenAI: researchers in the survey valued current coding agents above a large share of their team's staff. The 2 percent share of total research compute indicates that, by OpenAI's own measure, agent use is still a small part of the company's overall compute.
The task breakdown matters more than the headline usage numbers. Success is highest on monitoring and question answering and lowest on evaluation environments and inference code, which are the harder engineering tasks. Interviewees also credited humans with the central ideas behind GPT-6 Astra and GPT-6.2 Astra, apart from some pretraining insights on GPT-6.2 Astra. The token-cost comparison to salaries is theoretical and excludes equity, so it measures usage intensity rather than actual spending. The survey figures come from small, non-representative samples, which OpenAI itself notes.
Related Articles
OpenAI tests ChatGPT dots with a small group of Plus users, rolls out to all Business Standard
OpenAI has started testing ChatGPT dots with a small number of ChatGPT Plus subscribers and is rolling it out to all ChatGPT Business Standard users. Dots launched in September for Pro, Business Premium and Enterprise plans. Usage allowances for the new tiers have not yet been clarified.
OpenAI measures how coding agents accelerate AI research and where their performance stays uneven
OpenAI has published "Research acceleration: The jagged frontier," an analysis of how coding agents are speeding up AI research. It also covers where their performance remains uneven and what that signals about recursive self-improvement.
OpenAI models optimized circuits used in Jalapeño silicon and improved RISC-V core designs
OpenAI reports that its models optimized circuits used in Jalapeño silicon and improved RISC-V core designs. The company presents the work as progress in AI-assisted chip design with GPT-6 Astra.
OpenAI Python SDK v3.56.0 adds gpt-live-1.1 to Live model choices
OpenAI's Python SDK v3.56.0, dated 2026-11-19, adds gpt-live-1.1 to the Live model choices. The release lists this as its only feature.
Comments
Loading...