OpenAI says AI agents now finish nearly half of the multi-hour research tasks its own scientists hand them, without any human help. In a new report published on 18 November 2026, “Research acceleration: The jagged frontier,” the company shares internal data on how coding agents are changing the way it builds AI models, and argues that every frontier lab should have to publish indicators of recursive self-improvement.
Quick Take
- Autonomy: estimated agent success on tasks needing two to four hours of human work rose from 23% in March to 46% in October; on four to eight hour tasks it rose from 13% to 35%.
- Spend: at API list prices, the median researcher used about $2,400 of agentic AI per day in October, up 62% from September. The 90th percentile reached about $23,000, up 105%.
- Jagged: agents succeeded on 79% of attempts to monitor training runs but only 25% when writing code for inference and serving stacks.
- Safety staffing: documented work on Safety and Alignment grew 77% from January to September, the largest absolute increase of any research area.
- Value: in a survey of 101 research staff, the median answer was to give up 40% of team headcount or 30% of team compute to keep frontier coding agents for three months.
In plain English
OpenAI’s researchers now delegate a lot of their day to AI agents that write code, launch experiments and watch training jobs. This report measures how much of that work the agents can finish alone. The answer: a lot more than in spring, but very unevenly. Agents are good at some jobs and still weak at others, which is why OpenAI calls the frontier “jagged.” Humans still decide what to research and check the results.
What happened
The report updates a first research acceleration study OpenAI published in September. It covers three things: the inputs to AI research (mainly coding agents), how the research process is changing, and whether model capabilities are speeding up as a result.
OpenAI says it is sharing the data because, as it wrote in its frontier policy blueprint, “AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI).” It adds that “our reporting should not be read as an intention to pursue full RSI.” The topic has drawn attention since METR’s report on Anthropic, after which lawmakers called for a ban on recursive self-improvement.
What’s in the report
- Agent workdays: counting every eight hours of agent use as one workday, agent runtime per researcher rose from 0.5 to 5.7 workdays per day, an elevenfold increase. Pretraining and posttraining teams averaged roughly eight.
- Concurrency: by mid-October more than two-thirds of included researchers averaged a daily peak of at least four agents running at once, and more than 10% at least sixteen.
- Cost versus people: as of September, the API-equivalent value of agent use in the research organization was about twice its non-equity compensation. Internal deployment still accounts for only 2% of total research compute.
- Task categories: from 1 August to 18 October, estimated zero-intervention success was 69% for communicating, 58% for analysis and 43% for building. Building evaluation, reinforcement learning and agent environments came in at 29%.
- Researcher survey: teams said they would need a median of three months to match one month of current output without coding agents. In a separate October survey, 27 of 28 respondents chose or leaned toward the latest internal model over a mid-career human research engineer paired with a 2024 model.
- Safety shift: OpenAI says growth in Safety and Alignment work picked up after the Hugging Face incident in July 2026.
Why it matters
For developers, the task breakdown is a practical map. Monitoring, analysis and answering technical questions are where agents already pull their weight. Serving code and environment building still need close human review.
For companies, the spending figures show what heavy agent use looks like at a frontier lab: thousands of dollars per person per day at list prices, and agent costs that already exceed cash salaries.
For policy and safety, this is one of the few public data sets from inside a lab on how fast AI is speeding up its own development. OpenAI says the uneven progress means lower-performing task types “could become bottlenecks to high overall uplift.”
What’s still open
- The figures are self-reported by OpenAI and rely on internal classifiers and surveys that OpenAI itself calls partial.
- Whether overall capability progress is accelerating; OpenAI says it will keep tracking its capability index to find out.
- Whether other labs adopt shared reporting standards. OpenAI notes that Anthropic and Google DeepMind have published related measurements.
Sources
- OpenAI, “Research acceleration: The jagged frontier,” research report, 18 November 2026
- OpenAI, first research acceleration report, September 2026
i10x Perspective
The headline number is 46%, but the more useful finding is the spread. An agent that monitors training runs well but writes serving code poorly is a great assistant and a poor replacement. That is good news for teams adopting agents today: pick the tasks where success rates are high and keep humans on the rest. More from OpenAI and our Agents section.






