OpenAI research acceleration report: 46% of 2-4 hour tasks
OpenAI published its second research acceleration report on November 18, 2026 (US time). Its headline estimate is that coding agents completed 46% of internal research tasks sized at two to four hours of human work with no human intervention, up from 23% in March 2026. For four-to-eight-hour tasks the estimate rose from 13% to 35%. Results vary sharply by kind of work, from 79% for monitoring training runs to 25% for writing inference and serving code. OpenAI says the numbers are estimates graded by its own model, that the method changed since its September report, and that researchers may have changed which tasks they delegate. It writes that humans remain central to AI development
The three lines
- What rose — estimated zero-intervention success on two-to-four-hour tasks went from 23% in March 2026 to 46% in October 1-18, and from 13% to 35% on four-to-eight-hour tasks. Tasks under 15 minutes barely moved, from 77% to 82%
- What is jagged — for August 1 to October 18, 2026, agents succeeded on 79% of training-run monitoring and 72% of technical questions, but 29% of evaluation and RL environment builds and 25% of inference and serving code. OpenAI says the weak areas could bottleneck overall uplift
- What it does not say — a GPT-6 Astra agent judged both success and human time, with no outside verification. The $2,400 a day is usage valued at API list prices, not cost. The report contains no direct measure of research itself getting faster
Key questions
- openai research acceleration report jagged frontier
- **It is OpenAI's second self-measured report on how much its own researchers use coding agents and how often those agents finish work alone, published on November 18, 2026 (US time) under the title Research acceleration: The jagged frontier.** | Item | Detail | |---|---| | Published | November 18, 2026 (US time) | | Previous report | First report, September 2026 | | Zero-intervention success, 2-4 hour tasks | 23% (March) to 46% (October 1-18) | | Zero-intervention success, 4-8 hour tasks | 13% to 35% | | Strongest task | Monitoring training runs, 79% | | Weakest task | Writing inference and serving code, 25% | | Median researcher daily usage | About $2,400 at API list prices (October) | | 90th-percentile daily usage | About $23,000 | | OpenAI's conclusion | Uplift is jagged and humans remain central |
- what does openai 46% agent success rate mean
- **Of the tasks researchers actually handed to agents, 46% of those estimated to need two to four hours of human work were judged successful with no human stepping in. It is not a score on a fixed benchmark.** | Question | What the report discloses | |---|---| | Who graded | An analysis agent built on GPT-6 Astra | | Who measured human time | The same agent, by estimate. Not timed | | Uncertainty on the 46% | Approximate 95% interval of 33.0% to 58.5% | | Which tasks | Tasks researchers chose to delegate. The mix changes month to month | | Same method as September | No. Grader model, sampling, time estimate and handling of uncertain outcomes all changed | | Outside verification | None |
- how much do openai researchers spend on ai agents per day
- **As of October 2026 the median researcher used about $2,400 a day and the 90th-percentile researcher about $23,000 a day. These are tokens valued at public API list prices, not money OpenAI paid.** | Group | September 2026 | October 2026 | Increase | |---|---|---|---| | Median researcher | About $1,500 | About $2,400 | 62% | | 90th-percentile researcher | About $11,000 | About $23,000 | 105% | OpenAI notes that its researchers have no agentic coding quotas. A footnote adds that a temporary efficiency change deployed on August 21, 2026 depressed recorded spend, and that spend rose quickly once it was rolled back starting September 10.
OpenAI has put numbers on how often AI agents finish research work on their own inside its own labs. In a report published on November 18, 2026 (US time), "Research acceleration: The jagged frontier," the company estimates that coding agents completed 46% of internal research tasks sized at two to four hours of human work with no human intervention. In March 2026 the figure was 23%. For four-to-eight-hour tasks it rose from 13% to 35%. This is the second report in a series that began in September 2026.
The "jagged" in the title is the actual finding. Over the same period, agents succeeded 79% of the time at monitoring training runs and 25% of the time at writing inference and serving code. OpenAI's opening summary says humans are still central to AI development. This article covers what the report measured, what the numbers look like when split by task length and by task type, what they do not tell you, and how they relate to the recursive self-improvement debate and to METR's time horizon.
1. What kind of report this is — a lab measuring itself
OpenAI organizes the report into inputs, process and outputs. Inputs are how much researchers use agents. Process is what they delegate and how often it works. Outputs are whether code, experiments and model capability have grown as a result.
It also states why it is publishing. OpenAI's position is that AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI), the process by which AI accelerates the development of future AI systems, and it presents this report as doing so. In the same passage it adds that "our reporting should not be read as an intention to pursue full RSI."
| Section | Indicator | Figure in the report |
|---|---|---|
| Inputs | Median researcher's daily agent usage, at API list prices | About $2,400 in October 2026, about $1,500 in September |
| Inputs | 90th-percentile researcher's daily usage | About $23,000, versus about $11,000 in September |
| Inputs | Agent workdays per researcher per day (8 hours = 1 workday) | From 0.5 to 5.7 |
| Inputs | Concurrent agents | By mid-October 2026, more than two-thirds averaged a daily peak of four or more, and more than 10% sixteen or more |
| Process | Zero-intervention success, 2-4 hour tasks | 23% to 46% |
| Process | Zero-intervention success, 4-8 hour tasks | 13% to 35% |
| Outputs | Experiments per active experimenter | 1.68 times the 2025 baseline, averaged over the four weeks ending October 18, 2026 |
| Outputs | Rise in OpenAI's internal capabilities index (OpenAI ECI) | 16.4 points a year since August 2025, against an industry trend of 14 points a year observed by Epoch AI |
The dollar figures are not money spent. They are tokens used internally, valued at the list prices outside customers pay. OpenAI says that as of September 2026 this API-equivalent value was about twice the research organization's non-equity compensation. It also notes that equity, which it says generally makes up 70 to 80% of researchers' target compensation, is left out of that comparison, and that internal agent use accounts for only 2% of total research compute.
2. Inside the numbers — up by length, split by type
The chart data on OpenAI's page gives the following by task length. Success means a task judged successful with zero recorded human interventions. The range in parentheses runs from treating every uncertain outcome as a failure to treating every uncertain, zero-intervention outcome as a success.
| Estimated human time | March 2026 | October 1-18, 2026 | Share of October sample |
|---|---|---|---|
| Under 15 minutes | 76.5% (71.6-81.5) | 82.0% (76.9-87.1) | 19.5% |
| 15-30 minutes | 62.7% (57.4-68.0) | 77.8% (74.3-81.3) | 17.2% |
| 30 minutes-1 hour | 46.9% (40.8-52.9) | 66.9% (61.5-72.2) | 21.3% |
| 1-2 hours | 33.7% (27.5-39.8) | 54.8% (49.6-60.1) | 14.3% |
| 2-4 hours | 23.0% (17.9-28.2) | 45.7% (40.7-50.6) | 14.9% |
| 4-8 hours | 12.8% (7.6-18.1) | 35.1% (29.9-40.3) | 7.8% |
| 8 hours or more | 9.6% (6.8-12.5) | 29.2% (28.1-30.4) | 5.0% |
Three things stand out. First, short tasks hardly moved and long tasks moved a lot. Second, the mix of delegated work changed. Two-to-four-hour tasks went from 7.8% of the sample in March to 14.9% in October, and four-to-eight-hour tasks from 2.5% to 7.8%. Researchers are handing over longer work more often. Third, 46% read the other way means that more than half of two-to-four-hour tasks still involved a human stepping in, or failed. With sampling error included, the approximate 95% interval is wide: 33.0% to 58.5%.
Split by type of work, the picture changes. Using a taxonomy of AI R&D tasks developed by Epoch AI, OpenAI reports success rates for August 1 through October 18, 2026.
| Level | Task | Estimated zero-intervention success |
|---|---|---|
| Task | Monitoring training runs for issues | 79% |
| Task | Answering technical questions | 72% |
| Category | Communicate | 69% |
| Category | Analyze | 58% |
| Category | Build | 43% |
| Task | Launching training runs | 35% |
| Task | Building evaluation, reinforcement-learning and agent environments | 29% |
| Task | Writing code for inference and serving stacks | 25% |
Agents are good at watching and answering. At building, they succeed alone somewhere between one time in four and one time in three. OpenAI writes that as long as performance is uneven, the lower-performing tasks could become bottlenecks to overall uplift. A research cycle does not run on monitoring and Q&A alone.
Interviews with ten researchers point the same way. They report large gains in implementation, debugging and scanning thousands of samples, and they say agents have poor taste in research ideas, code quality and data quality, and that work goes wrong without a human close by. For GPT-6 Astra, researchers said that to their knowledge every key idea they named came from people. For GPT-6.2 Astra, pretraining researchers described developing ideas in conversation with Codex, and one credited it with identifying key insights.
3. What these numbers do not say
Most of the caveats are OpenAI's own. They are scattered through the text and the collapsed methods notes, so they are gathered here.
| Caveat | What the report discloses | What it means |
|---|---|---|
| The grader is an AI | A GPT-6 Astra agent extracts tasks from session logs and judges success and the number of interventions | No human reviewed the outputs. The judgment rests on signals such as user feedback, whether code was merged and whether later work built on the result |
| Human time is also an AI estimate | The same agent estimates how long a person would take. OpenAI says more validation is needed and the exact number may not be accurate | The two-to-four-hour bucket is not a measured duration |
| The method changed | New grader model, new sampling approach, multiple tasks per window, a new way of estimating time, and uncertain outcomes now included instead of dropped | The figures cannot simply be joined to the September 2026 report. OpenAI says the new time estimates come out lower |
| Task selection shifts | As agents improve, researchers may delegate different kinds of tasks within each bucket | The rise mixes capability gains with better choices about what to hand over |
| It is a subsample | A small subsample of interactive use. Known automation patterns are excluded | This is not the whole of the research organization's work |
| Roster issue | Data for March to June 2026 uses a present-day staff roster, not a historical one | The early population differs slightly |
| Spend is list-price value | The trend reflects more activity, shifts to pricier models and harness changes | More dollars does not mean proportionally more work |
| The survey is hypothetical | 101 submissions, non-random recruitment, and acknowledged limits of self-reports | Trading 40% of headcount or 30% of compute is an answer to a thought experiment, not a decision |
One more footnote attaches to the spend figures. OpenAI deployed a temporary harness efficiency improvement internally on August 21, 2026, which sharply depressed recorded token spend. It was rolled back gradually starting September 10, after which spend climbed quickly. Some of the 62% and 105% month-over-month growth may therefore be a rebound. The report does not say how much.
The larger gap is on the output side. More lines of code and more experiments mean more activity, and OpenAI itself writes that these do not necessarily imply faster research progress. The 16.4-points-a-year capability trend comes from an in-house index that blends public and internal benchmarks. Nowhere does the report estimate how many months sooner a model shipped because of agents. StartupHub.ai made the same point in its coverage: busy is not autonomous.
4. Between the recursive self-improvement debate and METR's time horizon
Both sides of the RSI debate can cite this report. Those who see acceleration point to success on two-to-four-hour tasks doubling in seven months, to 5.7 agent workdays per researcher, and to a survey in which the median respondent would give up 40% of team headcount to keep frontier agents. The cautious side points, in the same document, to humans still being needed on more than half of longer tasks, to ideas still originating mostly with people, and to a measurement that moved along with the models.
OpenAI adds one thing of its own in the conclusion. Trends built on human-AI collaboration may not predict future AI-driven progress. It cites recent sudden breakthroughs in mathematics by frontier models and says it is monitoring for similarly rapid breakthroughs in deep learning. A smooth curve, in other words, carries no promise of staying smooth.
METR's time horizon is a related number, but a different one. Both express task difficulty in human time. They are built differently.
| Aspect | METR time horizon | OpenAI's success by task length |
|---|---|---|
| Tasks | A fixed suite of test tasks | Whatever researchers chose to delegate in real work |
| Human time | Set from humans attempting the tasks, among other inputs | Estimated by an AI |
| Scoring | Task-specific grading | An AI judging from session logs and context |
| Human involvement | Assumed absent | Only zero-intervention cases count as success |
| Same tasks each month | Yes | No |
| Who measures | An outside nonprofit | The company itself |
On OpenAI's chart the last bucket above 50% is one to two hours (54.8%), with two to four hours at 45.7%. In day-to-day internal work, then, the coin-flip point sits somewhere around two hours. METR's public data, covered in the November 20, 2026 article "What is an AI time horizon," puts one model's 50% horizon at about 12 hours. The gap cannot be used to say either figure is wrong: the models, tasks and scoring all differ. It does point the same way as the benchmark saturation problem covered in the November 10, 2026 article on AI coding agent success rates. Tidy test tasks and messy real work produce different numbers.
What remains open and unconfirmed
| Open question | Current status |
|---|---|
| Outside verification | None. The raw data is internal session logs and is not public |
| Accuracy of the AI grader | The report text gives no figure for agreement with human review |
| Accuracy of the human-time estimate | OpenAI says more validation is needed |
| Separating capability from task selection | Not separated |
| How much faster research actually got | Not measured. Listed by OpenAI as future reporting |
| Whether capabilities or safety research is accelerating more | Not measured. Listed as future reporting |
| Actual internal cost | Not disclosed. Only list-price value |
| Full-month October figures | The chart ends on October 18 |
| Comparison with other labs | Methods differ, so no direct comparison is possible |
There is one safety figure too. OpenAI says documented researcher workweeks on Safety and Alignment rose 77% from January to September 2026, the largest absolute increase of any research area, with growth picking up after its response to the Hugging Face incident in July 2026. That number is also an estimate, built by running classifiers over Slack posts, pull requests and internal documents.
In short, the report shows that the way work is done inside the lab is changing fast. It does not show how much that change has altered the pace of research. The 46% is a company's estimate of its own work, graded by its own model, and it should be read next to the 25%.
What recursive self-improvement is and why it is contested is laid out in the October 16, 2026 article "What is recursive self-improvement." The gap between benchmark scores and tasks actually completed is covered in the November 10, 2026 article "AI coding agent success rate," and OpenAI's contracts to bring outside evaluators in-house in the November 18, 2026 article "OpenAI embedded assessments." How AI capability is measured in human time, and where that method falls short, is in the November 20, 2026 article "What is an AI time horizon."
Sources
- OpenAI — Research acceleration: The jagged frontier (2026-11-18)
- AI Weekly — OpenAI says coding agents finish 46% of 2-4 hour research tasks (2026-11-19)
- StartupHub.ai — OpenAI agents do busywork, not research (2026-11-19)
- The Rundown AI — OpenAI says agents finish nearly half of tasks that take researchers 2 to 4 hours (2026-11-20)
- AI TLDR — OpenAI's jagged frontier report (2026-11-18)
- Blockchain.News — OpenAI Reports Uneven Progress in AI-Driven Research Acceleration (2026-11-19)
- CellCog — OpenAI's Research Report: Top Researchers Spend $23K a Day (2026-11-19)