OpenAI’s internal coding agents now finish 46% of research tasks that would take a person two to four hours, with no human stepping in. In March the figure was 23%. The company published the numbers on November 18 in “Research acceleration: The jagged frontier,” its second report on how AI is changing its own research work. Every figure below is OpenAI’s own estimate from internal data, not an outside audit.

The headline rate hides a wide spread. For tasks sized at four to eight hours, success rose from 13% to 35% over the same months. By task type, agents succeeded 79% of the time on monitoring training runs, 35% on launching them, and 25% on coding the stacks that serve models to users. Between August 1 and October 18, unaided success was 69% for communication tasks, 58% for analysis, and 43% for building, using a taxonomy developed by Epoch AI.

OpenAI flags its own caveat. Researchers may now hand agents different jobs than they did in March, so a higher success rate reflects both better models and changed delegation habits. The company’s conclusion: the gains are uneven, and people are still central to the work.

Usage is climbing fast, and it is concentrated. Priced at the API’s published rates, a median researcher now consumes about $2,400 of agent work daily, up 62% from $1,500 in September. The 90th percentile researcher consumes about $23,000, up 105% from $11,000. Agent workdays per researcher, with eight hours of agent runtime defined as one workday, rose from 0.5 to 5.7, an elevenfold jump. Pretraining and posttraining teams averaged roughly eight; scaling and infrastructure averaged roughly four.

Parallelism is spreading too. By mid-October, more than two-thirds of the researchers in the analysis typically ran four or more agents simultaneously at their busiest, and more than 10% typically ran sixteen or more. 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 total research compute.

Researchers say they value the tools. In a survey of 101 research staff in late August and early September, the median respondent would give up 40% of team headcount, or 30% of team compute, to retain frontier coding agents over a three-month stretch. Asked the same question as of January 2027, the median compute figure was 50%. OpenAI acknowledges the sampling was ad hoc and self-reports can be inaccurate.

Interviews with ten researchers show where the tools fall short. They described agents as weak at judging what is worth researching, and at the quality of code, data, and prose, and said close supervision is needed to avoid “AI slop” pull requests. On the origin of big ideas, posttraining and safety monitoring researchers said people produced every idea they named for GPT-6 Astra and GPT-6.2 Astra. For GPT-6.2 Astra, pretraining researchers described working through ideas in conversation with Codex.

Safety work is the area growing fastest. Documented workweeks in Safety and Alignment rose 77% from January to September, the largest absolute increase of any major research area, through new hires and researchers shifting over. OpenAI says growth picked up after its response to the Hugging Face incident in July 2026. That category spans evals for alignment, monitorability, and work on reward hacking.

On output, each active experimenter logged 1.68 times as many experiments as the 2025 baseline, measured across the month to October 18. OpenAI’s internal version of the Epoch Capabilities Index shows its frontier climbing 16.4 points annually from August 2025, against 14 points per year that Epoch observed since late 2024. OpenAI notes its models score higher on its own benchmark mixture than on Epoch’s.

OpenAI uses the report to argue that every lab ought to be obliged to measure and openly disclose signs of recursive self-improvement, while saying it does not intend to pursue it. The data is self-reported by the lab being measured, so treat the trend lines as a floor for disclosure, not proof.

A budget lesson sits in the spend data: at list prices the 90th percentile researcher costs nearly ten times the median, so any team sizing agent budgets should plan for a small group of heavy users driving most of the bill.

Source: OpenAI, “Research acceleration: The jagged frontier,” published November 18, 2026.