Research & Measurement
OpenAI reports growing research-agent use, with human steering still essential
Internal measurements show more experiments and substantial agent activity, while OpenAI cautions that these indicators do not directly measure scientific progress.
What happened
OpenAI published an internal research-workflow study. It reports 3.1 agent-workdays per human workday by mid-August, using an eight-hour reference. More than half of successful tasks estimated at four to eight human hours involved intervention. The company says experiment counts rose alongside agent adoption and increased available compute.
Why it matters for AI
Usage, experiment volume and completed research are different measures. Human review and other bottlenecks can limit the value of faster execution.
AIRunsThis View
Our interpretation: evaluate accepted results, reviewer effort and failures together. Token volume or parallel runtime alone cannot establish productivity or replace human decisions.
What this could mean for us
Use outcome-level evidence when evaluating AIRunsThis workflows; do not apply another organisation’s ratios to our operations.
Source and method
This item is based on a primary source. AIRunsThis separates the source’s factual announcement from its own interpretation. Vendor performance claims have not been independently reproduced unless stated.