OpenAI estimates its coding agents completed 46% of delegated research tasks judged to require two to four hours of human work without intervention, up from 23% in March. The company’s November 18 report puts success on tasks estimated at four to eight human hours at 35%, up from 13%. The latest measurement period covers October 1–18, 2026. Source
OpenAI also updated its methodology and says researchers may have changed which tasks they delegated within each duration category. The figures don’t isolate how much of the increase came from better models. And those hours describe estimated human effort, not how long an agent actually ran. Source
The job description matters
For August 1–October 18, OpenAI estimated 79% success without human intervention on monitoring training runs for issues, but only 25% on writing code for inference and serving stacks—the software that runs models and delivers their responses. For building software environments used to test models, train them through rewards or run agents, estimated success without human intervention was 29%. Being good at watching the machinery doesn’t make an agent equally good at building it. Source
OpenAI’s synthesis of interviews with ten researchers describes agents producing unusable data, making invalid assumptions in analyses and writing “AI slop” pull requests, or proposed code changes. Researchers still needed to oversee the work closely to prevent low-quality outputs and wasted effort. The slop, apparently, can originate inside the lab too. Source
One anonymous researcher working on model evaluations in posttraining—the refinement stage after initial training—described checking an example before letting an agent repeat the work: “Now I say, ‘Go make one. I'll review it and make edits. Then go do the rest.’” Source
Big usage numbers, carefully priced
Researchers are nevertheless using agents much more intensively, according to OpenAI. In October, median daily usage reached about $2,400 when valued at public API list prices—the published rates for accessing models through software. At the 90th percentile, it reached about $23,000. The figures, calculated as 28-day averages, were up 62% and 105%, respectively, from September. They’re price-tag conversions, not actual internal bills or dollar values attached to discoveries. Source
OpenAI says the dollar-equivalent trend reflects increased activity, moves to more expensive models and changes in model behavior and the software coordinating agents. It also says a temporary efficiency improvement appears to have depressed recorded spending; OpenAI began reversing that change on September 10 for reasons unrelated to spending. A steeper spending chart isn’t a productivity meter.
Comments