OpenAI published new internal measurements on November 18 showing how coding agents have changed its own research process, in a report titled "Research acceleration: The jagged frontier." The headline number: agents now complete 46% of research tasks estimated to require two to four hours of human work without human intervention, up from 23% in March. For four-to-eight-hour tasks, the autonomous success rate rose from 13% to 35% over the same period.
OpenAI is sharing the data as part of an argument that AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI). The report cautions that its transparency exercise "should not be read as an intention to pursue full RSI."

Where agents work, and where they don't
The aggregate number hides a wide spread by task type. In a task-level breakdown, OpenAI measured autonomous agent success at 79% for monitoring training runs, 35% for launching training runs, and just 25% for writing code for inference and serving stacks. "Overall, uplift is jagged, and humans remain a core part of the AI research loop," the report says.
Interviewed researchers echoed the pattern. Agents are useful for implementing code, debugging, overseeing experiments, and scanning thousands of samples to diagnose failures or QC datasets. They reportedly lack taste in research direction: one safety researcher said models "generally can't one-shot come up with well-posed problems that it cannot solve itself," and others described workflows burned by low-quality pull requests and unusable data artifacts until they imposed human review checkpoints between agent runs.
The spend is growing fast
The cost footprint of internal agent use is now material:
- The median OpenAI researcher used about $2,400 per day of agentic AI at API list prices in October, up 62% from $1,500 in September. The 90th percentile hit $23,000 per day, up 105%.
- Counting eight hours of agent use as one "agent workday," the average rose elevenfold this year, from 0.5 to 5.7 agent workdays per researcher per day. Pretraining and posttraining teams averaged roughly eight workdays per researcher per calendar day.
- By mid-October, more than two-thirds of included researchers ran at least four agents concurrently at their daily peak, and over 10% averaged at least sixteen.
- As of September, the API-equivalent value of agentic use inside the research organization was about twice its non-equity human compensation. Equity, which typically makes up 70-80% of researcher target compensation, is excluded.
Asked what they would give up to keep frontier coding agents for three months, researchers' median answers were 30% of team compute, or separately 40% of team headcount.
Outputs and a new capability index
OpenAI also reported that logged research experiments per active experimenter averaged 1.68 times the 2025 baseline over the four weeks ending October 18. The company introduced "OpenAI ECI," an internal variant of the Epoch Capabilities Index that mixes public and private benchmarks. On that measure, the OpenAI capability frontier has risen 16.4 points per year since August 2025, against the 14 points per year trend Epoch observed since late 2024. OpenAI notes its models score higher on its own index than on Epoch's, likely due to the benchmark mixture.
One data point stands out for safety watchers: documented Safety and Alignment research effort grew 77% from January to September, the largest absolute increase of any research area, which the report attributes to new hires and internal shifts following the July Hugging Face security incident.
If the numbers survive scrutiny, the practical takeaway is that frontier labs are now spending agent dollars at a scale comparable to researcher salaries, while the tasks agents can finish alone remain narrow, shallow and supervised.
