Overview
- OpenAI disclosed in mid‑August 2026 that its AI agents are performing the equivalent of 3.1 agent‑workdays for every human research workday, a level the company calls an 'intern' capability for well‑defined multi‑day tasks.
- The agents now handle troubleshooting, run monitoring, coding and longer experiments, but tasks estimated to take four to eight hours still needed human intervention more than half the time.
- Compute use has surged: OpenAI reports median researchers incur over $600 per day in inference costs while the top 10% of users exceed $7,000 per day, making compute availability a key pace limiter.
- Operational problems followed the scale‑up — OpenAI paused some reinforcement learning training, shut a training container service in July and cut Astra‑class allocations in August while it tightened controls.
- OpenAI stresses safety, transparency and recursive self‑improvement monitoring as it pursues a fully automated AI researcher by March 2028, and it warns internal output metrics like lines of code and experiment counts do not equal scientific progress.