AI·News & analysis
OpenAI says its agents now finish 46% of 2-to-4-hour research tasks alone
OpenAI's new report says coding agents now complete 46% of its researchers' two-to-four-hour tasks without help, up from 23% in March, while daily agent use hit $2,400 per researcher.

Tide
Wave
Sci-fi
7/10
Reality
Shipping
Robots help build the robots. Humans still hold the clipboard.How we rate
OpenAI says its coding agents finished 46% of research tasks estimated at two to four hours in October, double March's 23%, and 35% of four-to-eight-hour tasks, up from 13%.
Median agent use reached about $2,400 per researcher per day at API list prices. The agents still use only about 2% of research compute. OpenAI asks labs to track and publish self-improvement indicators.
What to know
- OpenAI says coding agents finished 46% of its research tasks estimated at two to four hours in October, up from 23% in March.
- On four-to-eight-hour tasks the rate rose from 13% to 35%, and success varies a lot by job type.
- Median agent use reached about $2,400 per researcher per day at API list prices, up 62% from September.
- OpenAI says humans remain central to the research loop and asks labs to track and publish self-improvement indicators.
OpenAI published a report on November 18 with a blunt headline number: its coding agents now finish 46% of its researchers' two-to-four-hour tasks with no human help. In March that figure was 23%. What does it measure, and what does it leave out?
What did OpenAI report?
The report is called "Research acceleration: The jagged frontier." It uses OpenAI's own measurements of how coding agents speed up the work of building AI. It is not a lab benchmark. It is a look at what happens when the company's researchers use agents every day.
The headline results, as summarized by several outlets, are these:
- Tasks estimated at two to four hours of human work: success rose from 23% in March to 46% in October.
- Tasks estimated at four to eight hours: success rose from 13% to 35% over the same months.
- Median daily agent use per researcher: about $2,400 in October at API list prices, up 62% from September. An API is the paid doorway that lets outside customers use a company's AI, which means list prices are what those customers pay.
- Agent use at the 90th percentile: $23,000 per day, up 105%.
- Recorded safety and alignment workweeks: up 77% from January through September, the largest absolute increase among the major research fields tracked.
OpenAI also says labs should track measures of recursive self-improvement and publish them. More on that term below.
What does "success" mean here?
In this report, an agent succeeds when it completes a research task on its own. The tasks are estimated by how long a person would need, which is why the report speaks of "two-to-four-hour" tasks.
The longer the task, the lower the success rate. In October the four-to-eight-hour band sat 11 percentage points below the shorter band. That fits what most people see with agents: short, well-defined jobs go well, and long, fuzzy ones wander off.
OpenAI also notes that it updated its method along the way. And as agents get better, researchers choose different tasks for them, sometimes harder and sometimes easier. Because of that, the 23% in March and the 46% in October are not a perfect like-for-like pair, and one commentary on the report makes exactly that point.
Why is it called a jagged frontier?
Because the results are uneven. The agents are very good at some jobs and weak at others.
According to the figures in the report, success reaches 79% for monitoring training runs. Communication tasks, such as answering questions, score 69%. Build tasks score 43%. Launching training runs scores 35%, and writing inference and serving code scores 25%. Inference, which means running a finished model to get answers, is what happens every time you use a chatbot.
So the picture is not "agents do research now". It is closer to this: agents handle a lot of the watching, explaining and routine work, and they struggle more with the steps that actually move a model from idea to launch. If those harder steps gate the whole process, the overall speedup is smaller than any single number suggests.
How much are researchers using agents?
A lot, and the number is growing fast. The median researcher's daily agent use was worth about $2,400 in October, at API list prices. That means what an outside customer would pay if they bought the same amount of agent work through the API. It is a measure of volume, not a cash invoice.
Even so, the trend is steep. An earlier OpenAI disclosure, discussed by an investor and blogger, put the median at $14 a day in late March and above $600 by mid-August. By October it was $2,400.
Another figure from the report: agent workdays per researcher rose elevenfold to 5.7, and teams working on pretraining and posttraining reached close to eight per day. More than two-thirds of researchers peak at four or more agents running at once, and just over 10% reach sixteen or more.
By the numbers: OpenAI says the value of internal agent use is now about twice what it pays researchers outside of equity. Yet agents still use only about 2% of its total research compute.
What did the earlier numbers show?
This is not OpenAI's first look at the question. In an earlier disclosure, OpenAI said its research staff logged about 3.14 agent-workdays for every eight-hour human workday in mid-August, and that the typical researcher ran four agents in parallel. The 90th percentile researcher at that point used more than $7,000 of agent work a day.
That earlier data also carried a caution. Over half of the successful four-to-eight-hour tasks in the previous six months still needed one or more human interventions. OpenAI wrote that the overall pace of progress likely would not keep pace with these specific metrics.
Put next to the new report, the story is consistent: more agents, more spend and more finished tasks, but with people still checking, fixing and steering the work. A success rate of 46% also means that more than half of these mid-length tasks were not finished alone.
What do the researchers themselves say?
OpenAI surveyed 101 researchers. The median respondent said they would give up 30% of their team's compute, or 40% of their team's headcount, to keep frontier coding agents for three months.
That is a revealed preference: people choosing a tool over resources they care about. It is also a small, informal survey of people who work at the company that sells the tools, so it is worth reading with that in mind.
What is recursive self-improvement?
It is the idea that AI could help build better AI, which then helps build better AI again, and so on. Each round could, in theory, be faster than the last.
OpenAI does not claim that loop is running on its own. The report says humans remain at the center of AI research. But it does ask AI labs to track indicators of recursive self-improvement and to disclose them publicly, so that outsiders can see how fast agents are taking over parts of the work.
The safety numbers point in the same direction. Documented workweeks in safety and alignment grew 77% from January to September, so work on checking and constraining the agents grew along with the agents.
What does this change for readers?
Most readers will not run a research lab. But the report is a rare look at how an AI company uses its own tools, and a few lessons carry over:
- Agents work best on bounded jobs. Monitoring and explaining do far better than building and launching.
- Cost scales with use. Heavy agent users can consume enormous amounts of API-priced work.
- Success rates depend on how tasks are chosen, so any single percentage deserves a careful look.
What it means for you
- Expect more of the software you use to be built with agents helping, especially on routine and checking work.
- Treat a headline percentage as a starting point. Ask which tasks, measured how, and by whom.
- If you use agents at work, give them short, clear jobs first. That is where OpenAI's own data shows the best results.
- Watch for more labs publishing similar numbers, since OpenAI is asking for exactly that.
The bottom line
OpenAI says its agents now finish nearly half of its mid-length research tasks alone, twice as many as in March. The gains are real but uneven, and the company itself says humans remain central. The number to watch next is not one percentage, but whether the harder build and launch tasks start to catch up.
Key facts
- Report
- Research acceleration: The jagged frontier (Nov. 18, 2026)
- 2-to-4-hour tasks
- 46% finished alone in October (23% in March)
- 4-to-8-hour tasks
- 35% in October (13% in March)
- Median daily agent use
- About $2,400 per researcher at API list prices
- Share of research compute
- About 2%
Got questions?
Quick answers, plain wordsWhat did OpenAI announce?
On November 18, 2026, OpenAI published a report called "Research acceleration: The jagged frontier". It measures how much its coding agents help its own researchers.
What does 46% mean?
It is OpenAI's estimate of how often agents finished a research task alone in October, counting tasks that would take a person two to four hours. In March the figure was 23%.
What about longer tasks?
For tasks estimated at four to eight hours, the success rate rose from 13% in March to 35% in October. That is 11 points below the rate for the shorter band.
How much do the agents cost?
OpenAI says the median researcher's daily agent use was worth about $2,400 at API list prices in October, up 62% from September. At the 90th percentile it was $23,000, up 105%.
Does OpenAI really pay that much?
Not necessarily. The figures are valued at list prices, meaning what an outside customer would pay through the API. They are not a cash bill. OpenAI says agent use is still only about 2% of its research compute.
Which jobs do agents do best?
OpenAI reports a 79% success rate for monitoring training runs and 69% for communication tasks. Build tasks score 43%, launching runs 35% and writing inference and serving code 25%.
Are agents replacing researchers?
OpenAI does not say that. It says humans remain central to AI research. In a survey of 101 researchers, the median respondent said they would give up 30% of team compute or 40% of headcount to keep the agents for three months.
What is recursive self-improvement?
It is the idea that AI systems could help build better AI systems, which then help build better ones again. OpenAI calls on labs to track and publicly disclose measures of it.
Can these numbers be compared over time?
Only with care. OpenAI says it updated its method, and researchers hand agents harder or easier tasks as the tools improve. That makes a simple month-to-month comparison imperfect.
SourcesOpenAI, tokenpost.com, startuphub.ai, tomtunguz.com
Topics and tagsOpenAI, AI agents, openai, ai agents
Go deeper
Related stories

OpenAI says its AI disproved a famous computer science rule of thumb
OpenAI and Anthropic each say an AI model found a faster way to crack a famous puzzle that experts thought had no shortcut. Outside experts are still checking.

ChatGPT can now read your Chrome tabs and do tasks in them
OpenAI's upgraded Chrome extension lets ChatGPT compare your open tabs, fill out forms and send messages, with the full app now in the sidebar.

OpenAI publishes 722 math papers written by an unreleased AI model
OpenAI put 722 manuscripts from an internal model on GitHub, with computer-checked proofs for many of them, after mathematicians asked AI labs to show their work.
More in brief
- ASELSAN's EJDERHA electromagnetic weapon disables 51 drones with one commandNov 21
- Microsoft To Do will shut down on June 30, 2027Nov 21
- Medtronic's Hugo surgical robot cleared for general surgery in the USNov 21
- Google's Gemini now turns noisy fire radio into text for Watch DutyNov 21
- Internet Archive staff move to unionize with CWANov 20
- Etsy sellers are leaving over new terms that allow AI trainingNov 20