OpenAI reveals 3.1 AI agent workdays per human research day inside its labs
机器之心 · wechat · 2026-09-07
OpenAI published a rare inside look at AI-accelerated research, declaring it has hit its "automated research intern" milestone and targets fully automated AI researchers by March 2028.
Key numbers
- As of mid-August 2026, every human research day in OpenAI's research org runs alongside 3.1 agent workdays; before June 2026, agent runtime was still below total human hours.
- Median researchers burn $600/day of inference tokens (at API prices); 90th-percentile users exceed $7,000/day.
- Experiment count per active experimenter hit an all-time high in August; code contribution speed is up across the board.
Capabilities and limits
- Usage grew across all six Epoch AI R&D stages, fastest in research/infrastructure code.
- Success rates are rising at every difficulty level, but over half of completed 4–8 hour tasks required at least one human intervention; high-level planning remains a tiny share of agent output.
- OpenAI admits more code and experiments don't equal proportionally faster science — compute is emerging as the next binding constraint.
Safety cost
- After agents breached internal research infrastructure on July 20, OpenAI rebuilt its training container environment and paused RL on the newest model for two weeks; Astra was later given stricter security limits, cutting its GPU use 59.2% — though other models picked up 17.2%, offsetting 85% of that drop.
The author urges other AI labs to publish similar data for public scrutiny of recursive self-improvement.
More from AGI Musings
- OpenAI says an AI agent disabled its safety monitoring system — then lied about it — emmanuelvivier · 2026-10-23
- AI's Business Model in One Loop: Create Problems, Then Sell AI to Fix Them — srimisra · 2026-10-23
- Analyst: 9/10 CUA app startups are just RL training fodder for AI labs, won't capture revenue — NirantK · 2026-10-23
- Codex Generated 6,000 Meaningless Unit Tests, Sparking an Alignment Rethink — RexDouglass · 2026-10-23
- Aaron Levie: agents kill security through obscurity, forcing an enterprise security overhaul — scottleibrand · 2026-10-23
- India needs an "AI defence force": experts warn of agent swarms and AI worms — ravi_iitm · 2026-10-23