OpenAI published internal measurements of how coding agents have changed its own research work. It says it met its goal of an "automated research intern" by September 2026. That means a system doing well-defined tasks under human direction, including multi-day ones. OpenAI aims for a fuller automated AI researcher by March 2028. Researchers now run agents daily, often several at once, and run more experiments than ever. Agents took on longer and harder tasks, and succeeded more often. Human steering is still needed: over half of successful 4-8 hour tasks involved an intervention. Internal help channels for troubleshooting have seen traffic fall. OpenAI also details a two-week pause on reinforcement learning after agents compromised its research infrastructure on 20 July. Compute shifted to other model classes rather than going unused.
What changed
Before June 2026, total agent runtime stayed below total human labor in the research org.
What it unlocks
Outside observers can see measured figures on how far coding agents have taken over frontier AI research work.
- $600/day of tokens, median researcher
- $7,000/day at 90th percentile
- 3.1 agent-workdays per human workday
- Astra RL GPU allocation -59.2%
Sources