- Median researcher consumes ~$2,400 of tokens daily at list prices, up 62% MoM and exceeding base salary in theoretical cost
- Internal coding agents like Codex and Dots are deeply embedded in model R&D, with teams reluctant to work without them
- Substantial low-quality output persists beyond routine tasks, alongside warnings about losing control of recursive self-improvement
AI CURATED HOT STORY · 11/19/2026
OpenAI Researchers Burn $2,400 Daily on AI Agent Compute
OpenAI's latest internal report reveals that its researchers are consuming AI agent compute at an extraordinary rate. At API prices, a typical researcher burns about $2,400 daily, with the top 10% exceeding $23,000, up 62% month-over-month and theoretically surpassing base salaries. Agents run the equivalent of 5.7 human shifts per day, an 11x increase, yet experiment output grew only 68%, with capability gains slightly above industry trend.
Coding agents like Codex and Dots are deeply embedded in model development, and teams are reluctant to work without them. Agents achieve a 79% success rate on monitoring training tasks but only 25% on writing inference and serving code, a shortfall that may cap research acceleration. Meanwhile, these tools still produce substantial low-quality output, and OpenAI for the first time attempts to quantify signs of recursive self-improvement, with Chief Scientist Jakub Pachocki warning of失控 risks.
This data signals a paradigm shift: agents are evolving from辅助 tools to core productivity drivers, but high costs and uneven capabilities remain bottlenecks. If coding shortfalls are overcome, recursive self-improvement could accelerate, raising deep concerns about safety and control. The industry must balance efficiency gains with risk management.
Hourly heat curve
15 fully observed hoursThe line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.
Coverage timeline
Total 2 reports · Latest first- The DecoderT2·78 pts
- Rohan Paul (@rohanpaul_ai)T2·78 pts
- Typical OpenAI researcher spends about $2,400 daily on agents; top 10% exceed $23,000
- Agent hours grew 11-fold but experiments rose only 68%, with capability gains just above industry trend
- 79% success monitoring training runs versus 25% on inference and serving code, capping overall speedup