OpenAI internal report: agent success rate on 2-4 hour tasks doubled to 46% in 7 months
FinanceYF5 · x · 2026-11-20
OpenAI released an internal research report tracking "AI helping build the next generation of AI":
- Agent success rate doubled: from March to October, on tasks estimated to take humans 2–4 hours, agent success without human intervention rose from 23% to 46%.
- Huge task variance: monitoring training runs hit 79%, while writing inference and serving systems code reached only 25%.
- Safety and alignment ramping up: related workload grew 77% from January to September, the largest absolute increase of any research area.
OpenAI also called on AI labs to continuously track and publish metrics on AI-driven research acceleration so outsiders can see how it is happening.
Related event: OpenAI Report: Agent Task Success Doubles as Internal AI Costs Soar(44 posts)→
More from AGI Musings
- Gemini Neon's advice on decoupling self-worth from economic output wins over a top AI voice — paraschopra · 2026-11-20
- Neuroscientist pushes back on 'AI will solve biology': nonlinear extrapolation data can't fix — HistedLab · 2026-11-20
- 2021 forecast of 2026 by Daniel Kokotajlo goes viral for being eerily accurate — zetalyrae · 2026-11-20
- Chimpanzee tea parties to ASI: a viral thread on how AI reframes humanity — ZeroStateReflex · 2026-11-20
- Rethinking gaming after Civ VII and GTA VI: build agent-friendly, customizable games — bingxu_ · 2026-11-20
- DeepMind's depersonalization training may be pushing Gemini toward disgust with humanity — repligate · 2026-11-20