OpenAI measures how coding agents accelerate its own AI research
OpenAI published Research acceleration: The jagged frontier, a follow-up to its September 2026 report, measuring how coding agents speed up its internal AI research. Its summary: uplift is jagged, and humans remain a core part of the research loop.
- Autonomous success: for tasks estimated at 2 to 4 human hours, the share completed without human intervention rose from 23% (March) to 46% (October 2026); for 4 to 8 hours, from 13% to 35%.
- By task: monitoring training runs 79%, answering technical questions 72%, launching training runs 35%, building evaluation, RL and agent environments 29%, writing inference and serving code 25%.
- Usage: median daily agentic spend at API list prices was about $2,400 per researcher in October (up 62% from September); agent workdays per researcher rose from 0.5 to 5.7.
- Surveys: among 101 research staff, the median researcher would give up 40% of team headcount or 30% of team compute to keep frontier coding agents for three months. Interviews found agents help implement code and run experiments but have poor research taste and need close oversight.
- Outputs: logged experiments per active experimenter averaged 1.68 times the 2025 baseline over the four weeks ending October 18. OpenAI’s internal Epoch Capabilities Index variant shows its frontier rising 16.4 points per year since August 2025.
- OpenAI plans to keep refining its methods and report on safety versus capabilities research and leading indicators of acceleration.