OpenAI says coding agents now finish 46% of multi-hour research tasks

A new OpenAI report on its own research shows agents doing far more unaided work since March, with uneven results by task and rising spend per researcher.

Illustration of a jagged mountain ridge made of bar charts, with small robots climbing some peaks and a human researcher at a desk watching several screens below. Image: AgenticFeed illustration

OpenAI says its coding agents now complete 46% of internal research tasks that would take a person two to four hours, with no human intervention, up from 23% in March. The figures come from a report the company published Wednesday on how AI is speeding up its own research.

For longer tasks estimated at four to eight hours of human work, the zero-intervention success rate rose from 13% to 35% between March and October, according to the report, titled "Research acceleration: The jagged frontier."

The title refers to how uneven those gains are. OpenAI said agents succeeded 79% of the time when monitoring training runs but only 25% of the time when writing code for inference and serving stacks. "Overall, uplift is jagged, and humans remain a core part of the AI research loop," the company wrote.

Spending per researcher keeps climbing

The report puts a dollar figure on agent use inside the research organization. At API list prices, the median researcher used about $2,400 a day of agentic AI tools in October, up 62% from $1,500 in September, OpenAI said. At the 90th percentile, daily use reached about $23,000, up 105% from $11,000.

OpenAI said that as of September the API-equivalent value of that agent use was about twice the research organization's non-equity compensation. Internal deployment still accounts for only 2% of total research compute, according to the report.

Researchers are also running more agents at once. By mid-October, more than two-thirds of researchers in the analysis averaged a daily peak of at least four concurrent agents, and more than 10% averaged at least sixteen, the report says.

What researchers say they would give up

From late August to early September, OpenAI surveyed 101 research staff about a hypothetical loss of access, for three months, to coding agents better than GPT-5.2. The median respondent said they would trade 30% of current team compute to keep them, or, in a separate question, 40% of team headcount.

Respondents also said their team would need a median of three months to match the value of their last month of work without coding agents. OpenAI flagged "potential issues with the accuracy of self-reports" and described its survey sampling as ad hoc.

The report also lists recurring problems. Researchers said agents struggled to launch experiments efficiently and went down unnecessary paths. They also described agents writing "AI slop" pull requests and making invalid analysis assumptions.

More effort on safety

OpenAI said Safety and Alignment saw the largest absolute increase in documented research effort of any area, with documented workweeks up 77% from January to September. The company attributed the rise to new hires and to existing researchers shifting toward safety work, and said growth picked up after its response to the Hugging Face incident in July 2026.

On research activity, logged experiments per active experimenter averaged 1.68 times the 2025 baseline over the four weeks ending Oct. 18, according to the report. OpenAI cautioned that more activity does not necessarily mean faster research progress.

Why OpenAI is publishing it

OpenAI frames the report as tracking indicators of recursive self-improvement, the process by which AI speeds up the development of future AI systems. The company said AI labs "should be required to track and publicly report" such indicators, and added that its reporting "should not be read as an intention to pursue full RSI."

The report points to similar work elsewhere, including three Anthropic reports on internal research acceleration and an acceleration report that METR produced in collaboration with Anthropic.

All figures are OpenAI's own internal measurements and have not been independently verified. The company also noted that as agents improved, researchers may have changed which tasks they delegated in each duration bucket, which affects how the success rates compare over time.

Source: OpenAI, “Research acceleration: The jagged frontier”

Get the AgenticFeed newsletter

A free, regular roundup of the agent stories that matter. No spam, and you can unsubscribe at any time.

Want to reach readers following AI agents? Advertise with AgenticFeed

Next event · 19 Nov, 2026

MCP Dev Summit Nairobi

The Linux Foundation’s Nov. 19 and 20 summit brings Model Context Protocol developers to Kenya, with keynotes from OpenAI, GitHub and the Gates Foundation.

The AgenticFeed newsletter

A free, regular roundup of the agent stories that matter. No spam.