OpenAI says its coding agents now finish 46% of 2–4 hour research tasks unaided, up from 23%
OpenAI's second report on its own researchers. Agents still have poor taste and need humans in the loop, OpenAI says.
On 18 November 2026 OpenAI published its second internal report on how coding agents (AI programs that write and run code on their own) speed up its AI research. It tracks how often an agent finishes a task with no human stepping in. For tasks a person would take 2 to 4 hours, that share rose from 23% to 46% between March and October. For 4 to 8 hour tasks it rose from 13% to 35%.
Success varies by job. From 1 August to 18 October, agents handled 79% of “monitoring runs for issues” without help, 35% of launching training runs and 25% of writing code for inference and serving systems (the software that runs a finished model).
Use has grown fast. The median researcher’s agent use cost about $2,400 a day at API list prices in October, up from $1,500 in September. The busiest 10% spent about $23,000 a day. Researchers ran 5.7 agent workdays per day, against 0.5 before. In a survey of 101 staff, the median would give up 40% of team headcount rather than lose frontier coding agents for three months. Logged experiments per researcher were 1.68× the 2025 level.
OpenAI says agents have “poor taste in research ideas”, and that humans stay a core part of the loop. It adds that research taste may be the bottleneck for AI improving itself, and that labs should be required to report indicators of that.
What it means
A lab reports that agents now do a growing share of multi-hour research tasks on their own, and that its staff value them highly. The figures describe OpenAI’s own work. They do not show how much faster its models improve, and OpenAI does not claim they do.
Caveats
- Everything is self-reported by OpenAI, with no outside audit. METR’s audit of Anthropic found AI R&D about 1.9× faster, but it used a different measure and an outside reviewer.
- The success rates may partly reflect which tasks researchers choose to hand to agents, not only what agents can do.
- The staff survey relies on self-reports, and OpenAI describes its sampling as ad hoc.
- A temporary change to the agent software in August lowered recorded spend until it was reversed from 10 September, so the September-to-October rise overstates the underlying trend.