On Wednesday OpenAI published its second report on how much coding agents are speeding up its own AI research. It follows a first report from September. The title, “The jagged frontier,” sums up the finding: the agents are getting better fast, but very unevenly.
Success rates by kind of task
For tasks that would take a human two to four hours, the share solved with no human intervention rose from 23 percent in March to 46 percent in October. For four-to-eight-hour tasks it went from 13 to 35 percent. The type of work matters a lot. Agents handle monitoring training runs on their own 79 percent of the time, launching training runs 35 percent of the time, and writing code for inference and serving stacks 25 percent of the time.
What the agents cost
At API list prices, the median researcher now uses about $2,400 of agent time per day, up 62 percent from September. At the 90th percentile it’s $23,000. Counted in eight-hour agent workdays, each researcher now gets 5.7 per calendar day, up from 0.5 at the start of the year. More than two thirds of the researchers tracked run at least four agents in parallel at their daily peak.
In September, the API-equivalent value of agent use was about twice the research organization’s pay, not counting equity. Internal use still accounts for only 2 percent of total research compute. In a survey, researchers were asked what they would give up to keep frontier coding agents for three months. The median answers: 30 percent of their team’s compute, or 40 percent of its headcount.
Most of the extra human effort went into safety and alignment, where documented work has almost doubled since January. OpenAI dates the bend in that curve to the weeks after the Hugging Face incident in July, the one California is now suing OpenAI over.
Measuring what it asked others to measure
In September, OpenAI called for AI labs to publicly report signs of recursive self-improvement. This report is OpenAI following its own advice. A footnote credits Anthropic, which has already published three such reports, and a METR study on Claude. OpenAI’s numbers and Anthropic’s point the same way: people are handing more and more research work to models, and the labs now track for themselves how fast that is happening.