OpenAI on November 18, 2026 published an update on research acceleration, based on internal measurements of how coding agents have changed its AI development cycle. The headline result is that agents now complete nearly half of two-to-four-hour research tasks on their own, although the company stresses that humans remain central to the work and that the uplift is uneven.

Between March and October, estimated success without human intervention rose from 23% to 46% for tasks that would take a person two to four hours, and from 13% to 35% for tasks estimated at four to eight hours. The type of work matters a great deal. For August 1 to October 18, OpenAI reports success rates of 69% for communicating, 58% for analyzing and 43% for building. At task level the range runs from 79% for monitoring training runs and 72% for answering technical questions down to 35% for launching training runs, 29% for building evaluation, reinforcement-learning and agent environments, and 25% for writing code for inference and serving stacks. OpenAI notes that weaker task types could become bottlenecks for overall uplift, and that changes in success rates can also reflect researchers changing what they delegate.

Usage is growing quickly. At API list prices, the median researcher now uses about $2,400 of agent tools per day, up 62% from $1,500 in September, while the 90th percentile reached about $23,000, up 105% from $11,000. Agent workdays per researcher, counting every eight hours of agent use as one, rose from 0.5 to 5.7, with pretraining and posttraining teams averaging roughly eight. By mid-October more than two-thirds of included researchers peaked at four or more concurrent agents, and more than 10% averaged at least sixteen.

OpenAI also looked at how researchers value the tools. In a survey of 101 research staff, the median respondent said they would give up 40% of team headcount or 30% of team compute to keep access to frontier coding agents for three months, and 50% of compute if asked in January 2027. As of September, the API-equivalent value of agent use was about twice researchers' non-equity compensation, yet internal deployment accounts for only 2% of research compute. On ideas, researchers behind GPT-6 Astra and GPT-6.2 Astra told OpenAI that people originated the ideas they named and models helped implement them.

The report also tracks where effort goes. Documented workweeks for Safety and Alignment, covering work such as alignment evaluations, monitorability and reward-hacking research, rose 77% from January to September, the largest absolute increase of any research area, with growth picking up after OpenAI's response to the Hugging Face incident in July.

OpenAI says it shares these figures because AI labs should be required to track and publicly report indicators of recursive self-improvement, and it states that the reporting does not signal an intention to pursue full recursive self-improvement. It points to related measurements from Anthropic, Google DeepMind and METR.

The caveats are significant. All figures come from OpenAI's internal data and surveys, which it calls a partial picture. A harness change on August 21 temporarily depressed recorded token spend before being reverted from September 10, and the survey sampling was ad hoc. No outside party has verified the numbers.

References