Skip to content
OpenAI News

OpenAI research acceleration report: agent capability is uneven across tasks

Research acceleration: The jagged frontier

OpenAI published an internal research acceleration report. Coding agents now complete tasks expected to take a human 2 to 4 hours 46% of the time, up from 23% in March. Results vary by task: monitoring training runs succeeds 79% of the time, writing inference and serving stack code 25%. Researchers still need to stay closely involved in supervision and quality control.

Why it matters: OpenAI breaks out agent autonomy by task type, and the gap between monitoring training runs and writing inference serving code shows the uneven payoff of research automation.

Read the original ↗Export Markdown