OpenAI's own researchers now say they value their AI tools more than nearly half their colleagues
Reported by The Decoder1 source ↓
OpenAI reports that internal coding-agent use has risen, with agents’ theoretical API value reaching twice researchers’ salary costs in September, excluding equity. Researchers also described errors and poor-quality code that require close human oversight.
Key points
- The median researcher’s agent use had a theoretical API list-price value of $2,400 per day in October, up 62% from September
- For the top 10% of users, the theoretical daily value reached $23,000
- Agents ran the equivalent of 5.7 eight-hour workdays per researcher per calendar day
- In a poll of 28 employees, 27 preferred the latest internal model to a mid-career human engineer
- OpenAI’s report says agents do better on routine tasks than complex engineering work
Details
OpenAI says internal use of tools including Codex and Dots has increased. At regular API list prices, median usage in October was equivalent to $2,400 in tokens per day; for the top 10% of users, it was $23,000. The company estimated that agents ran for the equivalent of 5.7 eight-hour workdays per researcher per calendar day, and about eight in pre- and post-training teams.
OpenAI said the theoretical API value of agent runtime was twice researchers’ salary costs in September. The calculation excludes equity packages, which the report says make up 70% to 80% of total compensation, and agent use accounted for 2% of the company’s total research compute. More than two-thirds of tracked researchers ran at least four agents in parallel at peak usage.
The company estimated that autonomous agents’ success on tasks lasting two to four hours increased from 23% in March to 46% in October; on four- to eight-hour tasks, it rose from 13% to 35%. Agents completed training-run monitoring 79% of the time, but succeeded on building evaluation and reinforcement-learning environments 29% of the time and writing inference and serving code 25% of the time.
In a non-representative survey of 101 employees, the median respondent said they would rather lose 40% of their team’s staff or 30% of its compute capacity than go three months without current coding agents. Interviews with ten researchers also surfaced problems including incorrect assumptions, unnecessary steps and low-quality code and data artifacts. OpenAI says the data may not extrapolate linearly.
Written from coverage as of