OpenAI Says Coding Agents Now Finish 46% of Two-to-Four-Hour Research Tasks on Their Own

OpenAI coding agents finished 46% of two-to-four-hour research tasks on their own in October, up from 23% in March, the company said, with uneven results.

Share this article

Featured image: OpenAI’s headquarters at 1515 Third Street in San Francisco’s Mission Bay, in a June 15, 2025, file photo. The image is not from the Nov. 18, 2026, research report. Photo: Coolcaesar, Wikimedia Commons, CC BY 4.0. Cropped to 16:9.

OpenAI said Wednesday, Nov. 18, 2026, that its OpenAI coding agents now complete 46% of research tasks estimated to take a person two to four hours, up from 23% in March, and 35% of four-to-eight-hour tasks, up from 13%. The company called the gains “jagged,” meaning they vary widely by the kind of work, and said people remain central to its research.

Advertisement
2–4 hour tasks
46%
autonomous success in October, up from 23% in March, OpenAI said
4–8 hour tasks
35%
up from 13% over the same period, OpenAI said
Median daily agent use
$2,400
per researcher in October, up 62% from September, at API list prices
90th percentile
$23,000
per researcher per day in October, up 105% from September, OpenAI said

The figures come from OpenAI’s own internal measurements, published in a report titled “Research acceleration: The jagged frontier.” They have not been independently audited. OpenAI said it is sharing them because labs should have to track and publicly report indicators of recursive self-improvement, or RSI — AI systems speeding up the development of later AI systems — and that the report should not be read as a plan to pursue full RSI.

Success is uneven by task

In a task-level breakdown, OpenAI said the autonomous success rate was 79% for monitoring training runs, 35% for launching training runs, and 25% for writing code for inference and serving stacks. The company said that pattern is why it describes the uplift as jagged: agents do far better at some jobs than others, and people still sit in the research loop.

The 2-to-4-hour success rate doubled, from 23% to 46%, a Boston Report calculation from OpenAI’s two figures. The 4-to-8-hour rate rose 22 percentage points, from 13% to 35%, more than doubling as well. Those comparisons use OpenAI’s own estimates of how long a person would take, and OpenAI did not publish the underlying task list.

Grouped bar chart of OpenAI’s estimated autonomous agent success rates: 23 percent in March and 46 percent in October on two-to-four-hour research tasks, and 13 percent in March and 35 percent in October on four-to-eight-hour tasks
Source: OpenAI, “Research acceleration: The jagged frontier,” Nov. 18, 2026. Chart: The Boston Report.

Spending, surveys and a shift toward safety work

Agent use inside the company is rising quickly, OpenAI said. At API list prices, the median researcher’s daily usage reached $2,400 in October, up 62% from September. At the 90th percentile it reached $23,000, up 105% from September. OpenAI noted that its researchers do not currently have agentic coding quotas, and that an internal “harness” change in late August temporarily depressed measured token spending before it was rolled back starting Sept. 10.

A survey of researchers, OpenAI said, asked what they would trade to keep frontier coding agents for three months. The median answers were 30% of team compute or, in a separate question, 40% of team headcount. AI Magazine, writing on Nov. 20, reported additional comments from OpenAI chief economist Aaron Chatterji, who said the most consequential ideas behind GPT-6 Astra “came from people, with agents helping put those ideas into practice,” and that the company needs to “share data regularly to share our own understanding of trends and to allow people to draw their own conclusions.”

Safety and Alignment saw the largest absolute increase in documented research effort among OpenAI’s major research areas, the company said. Documented workweeks in that area rose 77% from January to September, OpenAI said, because of new hires and because existing researchers shifted toward that work. The company said growth picked up after its response to a July 2026 incident involving Hugging Face.

Advertisement

Why OpenAI is publishing this now

The report builds on a September update on research acceleration at the company. OpenAI said it wants common measurement norms across labs, and it pointed to work by Anthropic, Google DeepMind and METR on related questions. The same week, California released recommendations that would give the state more authority over RSI and emergency shutdowns, a debate covered in our AI news and AI policy reporting.

OpenAI’s numbers describe work inside one company, on tasks it defined, over a period when its own models were also changing, including the GPT-6.2 Astra and Sol release. They do not measure how much those agents have moved the wider research frontier, and OpenAI did not claim that they do.

Related: California Panel Urges Emergency Power to Shut Down Dangerous AI Systems; OpenAI Says Its Models Improved AMD’s Flagship GPU Design by Up to 14%.

Sources

Read next in Artificial Intelligence

Get The Boston Report Briefing

Our free weekly email with the most important stories from Boston and the world. Unsubscribe anytime.

Free. We never sell your address. Privacy policy

Get The Boston Report Briefing
Our free weekly email with the most important stories from Boston and the world. Unsubscribe anytime.
Sign up