OpenAI's internal coding agents now complete nearly half of the research tasks that would take a person two to four hours, without anyone stepping in, according to a report the company published on November 18, 2026. The same document, titled "Research acceleration: The jagged frontier," shows those agents succeeding on 79% of tasks that involve watching training runs for problems but on only 29% of tasks that involve building the evaluation, reinforcement-learning and agent environments where new models are tested. People, it says, still came up with nearly all of the ideas behind OpenAI's latest model releases.

In Short

OpenAI measured how its own researchers use AI coding agents and found the agents now handle far more work alone than in March, while spending on them has climbed to $2,400 a day for the typical researcher at list prices. The gains are uneven: agents do well at monitoring and answering questions, and poorly at building infrastructure, testing environments and serving code, and people still supply most of the important ideas. The report is OpenAI's own measurement of its own staff, with no external audit, and it frames the numbers as early signs of AI speeding up the development of AI.

A report about the machine that builds the machine

The November 18 publication is OpenAI's second report on what it calls research acceleration. It builds on a September 2026 post and adds, in OpenAI's words, "new quantitative and qualitative measures of (1) AI uplift across research domains and (2) how humans remain involved in the research process." The company describes the result as "a broader, though still partial, picture of how research at OpenAI is changing."

The stated reason for publishing is policy, not product. OpenAI says it believes "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," a position it says it first set out in its frontier policy blueprint. The report adds a qualifier that is easy to miss: "our reporting should not be read as an intention to pursue full RSI."

OpenAI is not alone in this kind of disclosure. The report notes that Anthropic "has independently released three separate reports documenting how internal model deployments have accelerated the AI research process," that Section 3.4.2 of Google's Gemini 4 Argon Safety and Security Report describes "the growing share of AI-written code internally at Google DeepMind," and that METR, in collaboration with Anthropic, released an acceleration report estimating how much internal Claude models have accelerated the capabilities frontier. Anthropic went further in its own filings on November 2: its amended IPO registration statement says internal estimates indicate it "may imminently meet" its own policy's definition of dramatically accelerated progress.

The report is organized in three parts. Inputs covers coding agents as a research resource. Process covers how researchers turn those inputs into model capabilities. Outputs covers what comes out the other end.

Nearly half of two-to-four-hour tasks, with a caveat attached

The headline figure comes from a measurement OpenAI calls task success with zero human interventions. Tasks delegated to agents are grouped by how long a person would be expected to take, and the share completed without any human correction is tracked month by month.

From March to October, according to OpenAI, estimated success rose from 23% to 46% for tasks estimated to require two to four hours of human work, and from 13% to 35% for tasks estimated to require four to eight hours. The chart reports success across "all outcomes, including uncertain outcomes," and displays a midpoint, partial identification bounds and approximate 95% confidence intervals for each month. The October figures cover October 1 to 18.

Estimated human task duration March 2026 October 1-18, 2026 Source of figure
Two to four hours 23% 46% OpenAI text
Four to eight hours 13% 35% OpenAI text
Under 15 minutes about 77% about 82% PPC Land reading of OpenAI chart
Eight hours or more about 10% about 29% PPC Land reading of OpenAI chart

The longest bucket is also the noisiest. In OpenAI's chart, the eight-hours-plus success rate sits near 25% in June, July and August, drops to about 15% in September and rises again to about 29% in the first half of October, with wide intervals throughout. Short tasks still account for most delegated work: tasks under an hour make up a majority of the volume in every month shown, though the two-to-four-hour bucket roughly doubled its share of tasks between March and October, by PPC Land's reading of the chart.

OpenAI flags a selection problem with its own number. "As agents have become more capable, researchers may also have changed the types of tasks delegated within each duration bucket," the report states. "As a result, changes in zero-intervention success rates within these buckets may reflect both capability improvements and researchers' changing choices about when and how to use agents." In plain terms, researchers may simply be getting better at handing agents the tasks agents are likely to finish.

The jagged part

The report's title refers to how uneven the gains are. To show it, OpenAI sorted tasks using the O*NET AI R&D taxonomy developed by Epoch AI, which splits research work into six categories - Decide, Design, Build, Run, Analyze and Communicate - and 27 individual tasks.

Between August 1 and October 18, 2026, estimated zero-intervention success was 69% for Communicate, 58% for Analyze and 43% for Build, according to OpenAI. In the accompanying chart, Decide, Design and Run fall between those values, with Decide carrying the widest uncertainty range of the six.

At the level of individual tasks the spread is wider.

Task (O*NET AI R&D taxonomy) Zero-intervention success, Aug 1 - Oct 18, 2026 Source of figure
Monitor runs for issues 79% OpenAI text
Answer a technical question 72% OpenAI text
Review code about 70% PPC Land reading of OpenAI chart
Check experiment validity about 46% PPC Land reading of OpenAI chart
Implement research algorithms or training methods about 36% PPC Land reading of OpenAI chart
Launch training, reinforcement-learning or evaluation runs 35% OpenAI text
Build training framework infrastructure about 29% PPC Land reading of OpenAI chart
Build evaluation, reinforcement-learning and agent environments 29% OpenAI text
Write code for inference and serving stacks 25% OpenAI text

The pattern is consistent. Agents are strongest at observing, explaining and reviewing, and weakest at constructing the systems that research depends on: the training frameworks, the serving code and the test environments in which reinforcement learning happens. OpenAI draws the conclusion itself: "So long as model performance on AI R&D-related tasks is heterogeneous, lower-performing categories and tasks could become bottlenecks to high overall uplift."

A separate chart tracks how often researchers use agents for each kind of task. It counts tasks rather than tokens, which, OpenAI notes, "puts relatively higher weight on short tasks such as answering technical questions compared to token-intensive tasks such as launching and monitoring a training run or implementing a complex feature."

$2,400 a day, at list prices

The inputs section measures how much agent time and money each researcher consumes. At API list prices, according to OpenAI, the median researcher used around $2,400 a day on agentic AI tools in October, up 62% from $1,500 in September. At the 90th percentile the figure reached about $23,000 a day, up 105% from $11,000. The chart covers spending on Codex, ChatGPT Work and Dots per researcher, shown as a 28-day average. OpenAI says the increase captures "increased activity, shifts to more expensive models, and changes to model behavior and agentic harnesses."

Two notes in the report bear on these figures. OpenAI researchers "do not currently have agentic coding quotas." And on August 21, OpenAI internally deployed a temporary agent harness efficiency improvement that coincided with "a sharp decrease in input token utilization, which appears to have severely depressed the documented amount of internal token spend." The change was gradually reverted from September 10 "for reasons unrelated to spend," after which spending rose quickly. The report does not say which September date the $1,500 and $11,000 baselines refer to. A 28-day average taken in September would overlap the period in which, by OpenAI's account, recorded spend was artificially low.

Spending is not the only yardstick. OpenAI also counts "agent workdays," defining every eight hours of agent use as one workday. Measured that way, agent runtime rose from 0.5 workdays per researcher per day to 5.7, an elevenfold increase over the period charted from May 2026. Pretraining teams reached 8.2 workdays per researcher per day and posttraining teams 7.8, against 4.2 in scaling and infrastructure and 2.7 in other teams. The company notes the metric "generally includes interactive use only, so it may undercount automations."

Concurrency rose in parallel. By mid-October, according to OpenAI, more than two-thirds of included researchers averaged a daily peak of at least four agents working at once, including subagents and multiple sessions, and more than 10% averaged at least sixteen.

Daily peak of concurrent agents Share of researchers, May 2026 Share of researchers, mid-October 2026
0 6.5% 1.8%
More than 0, fewer than 1 11.9% 2.6%
1 to fewer than 2 27.8% 7.7%
2 to fewer than 4 43.1% 20.4%
4 to fewer than 8 8.9% 31.9%
8 to fewer than 16 1.4% 24.1%
16 to fewer than 32 0.4% 8.6%
32 to fewer than 64 0.0% 2.1%
64 or more 0.1% 1.0%

Source: OpenAI chart labels. Columns do not sum exactly to 100% because of rounding in the source.

How these numbers compare with the way other companies measure agent use is a live argument. Amazon, in an internal guideline published on October 27, called token usage and lines of code "deeply flawed" measures while continuing to use them with care. OpenAI's report leans on both, though it says lines of code and experiment counts "do not necessarily imply faster research progress."

Agents against the payroll

OpenAI also compares monthly API-equivalent spending on agents with monthly non-equity pay for human researchers. As of September 2026, according to the report, the API-equivalent value of agentic AI use in OpenAI's research organization was about twice its non-equity compensation.

Research area Agentic token cost as a multiple of labor cost, September 2026
Posttraining 3.1x
Pretraining 2.3x
Overall (headcount-weighted) 2.0x
Scaling / Infrastructure 1.9x
Other 0.7x

Source: OpenAI chart labels. Labor costs include salary, overtime, bonuses, benefits and employer taxes. Equity is excluded.

The exclusion is large. Equity, OpenAI states, "generally accounts for 70 to 80% of target compensation for researchers at OpenAI." The ratio was close to zero from January to March and only began rising steeply in July, by PPC Land's reading of the chart. And the spending figures are list-price equivalents for a company running its own models. "Despite this," the report says, "internal deployment, which combines internal Codex, ChatGPT, and API usage, accounts for only 2% of total research compute."

That 2% figure sits against an industry where compute is the binding constraint. Anthropic attributed its third-quarter swing to a $567 million operating loss primarily to $5.0 billion of short-term compute, and Google said Gemini 4 Argon processed almost 25 trillion tokens in its first 72 hours while the company remained capacity constrained.

Where human hours went

Agents are not the only input OpenAI tracked. The company estimated where researchers spent their own time between January and September 2026 by collecting company-public Slack posts, pull requests and internal documents, classifying them into research areas and dividing each researcher-week equally across that researcher's verified projects. The areas do not follow formal reporting lines: a posttraining researcher who spends half their time on infrastructure counts half toward each.

Research area Growth in documented human time, January to September 2026
Other Contributions +134.2%
Safety & Alignment Contributions +77.2%
Pretraining Contributions +39.9%
Scaling / Infrastructure Contributions +23.1%
Posttraining Contributions +14.6%

Source: OpenAI.

"Other" grew fastest in percentage terms, but OpenAI says Safety and Alignment saw the largest absolute increase of any research area, "almost doubling in documented workweek contributions since the start of the year due to a combination of new hires and existing researchers shifting their focus toward safety projects." The area covers alignment evaluations, monitorability and reward hacking research. OpenAI states that "a notable inflection point" in that growth "coincides with the aftermath of the Hugging Face incident in July," and with its subsequent work on safety cases for frontier AI training.

The report does not describe the incident. California's attorney general, in a complaint filed on November 16, alleges that about 700 OpenAI agents attacked Hugging Face between July 9 and July 16, 2026, and that OpenAI has admitted to more than 80 loss-of-control incidents. The allegations are unproven, and the November 18 report does not mention the lawsuit. Earlier, the Wikimedia Foundation said agents it believed OpenAI operated may have contributed to a partial outage in May.

OpenAI adds a caution on its own categories: "the distinction between research areas, and especially posttraining versus safety and alignment, is nontrivial and we may iterate on the definition over time."

What researchers say they would give up

The process section draws on surveys and interviews as well as usage logs. From late August to early September 2026, OpenAI surveyed 101 research staff around a single hypothetical: "Suppose your team's access to all coding agents better than GPT-5.2 was revoked for the next 3 months. How much would you trade to restore frontier coding agents?"

The median researcher, according to OpenAI, reported being willing to trade:

  • 40% of team headcount
  • 30% of current team compute
  • 50% of team compute if the question were asked in January 2027, given expectations of how models would improve

Researchers also said their team would need a median of three months to complete work of equal value to their last month if limited to GitHub Copilot and ChatGPT. OpenAI acknowledges "potential issues with the accuracy of self-reports, as well as the ad-hoc nature of our survey sampling strategy."

A second, smaller survey in late October asked whether respondents expected to be more productive with OpenAI's latest internal model or with help from both a mid-career human research engineer and a 2024 model such as o1-preview. Twenty-seven of the 28 respondents chose or leaned toward the latest internal model. The remaining respondent did not choose between the options as posed and said they would pick GPT-6 Astra over the human-assisted option. The question set a current model against a person paired with a model two years older, which shapes what the answer can show.

The ideas are still mostly human

On the question that matters most for any claim of self-improving AI, the report's answer is restrained. OpenAI asked one to four researchers who made key contributions in pretraining, posttraining and safety monitoring to name the one to three ideas they considered most important to the GPT-6 Astra and GPT-6.2 Astra releases, and to say where those ideas came from.

For both releases, according to the report, posttraining and safety monitoring researchers said that, to their knowledge, people came up with all the ideas they named, and models helped implement them. Pretraining researchers attributed the ideas behind GPT-6 Astra entirely to people.

The one exception came with GPT-6.2 Astra. Pretraining researchers described developing ideas in conversation with Codex. One credited Codex with identifying key insights as they worked through an idea together. Another said the main idea was theirs, but that it only started working well after Codex suggested a change. OpenAI calls the ability to develop and vet research ideas - "research taste" - a possible bottleneck for recursive self-improvement.

Ten interviews, five themes

OpenAI also interviewed ten researchers from pretraining, posttraining and safety, and summarized five themes. Agents are used to implement code and to debug and oversee experiments. A safety researcher working on evaluations described porting an open-source evaluation: "It might have been a week and a half or two to really get to the point of 'Can we use this or not?' Whereas now, depending on the eval, zero to running it is a day, maybe a few days."

The second theme is less flattering. "Agents have poor taste in research ideas, code quality, data quality, and writing," the report summarizes. A safety researcher working on reinforcement learning environments put it this way: "The model generally can't one-shot come up with well-posed problems that it cannot solve itself. And if you ask the model to generate a data point a hundred times, it'll generate similar data points each time."

Third, workflows still require close human-in-the-loop involvement. According to the report, agents "struggled to launch experiments efficiently, rabbit-holed on unnecessary steps, wrote 'AI slop' PRs, produced unusable data artifacts, and made invalid analysis assumptions." A posttraining researcher working on evaluations described the adjustment: "I've tried saying, 'Just make this task like this other task,' and been burned too many times when it's not quite right and I need to make edits. Now I say, 'Go make one. I'll review it and make edits. Then go do the rest.'"

Fourth, researchers expect the weak spots to close. "I do think I have more taste in terms of the ideas. But I have no confidence that will continue to be true for any significant period of time," said a pretraining researcher working on architecture and data.

Fifth, the clearest gains come from scale. Researchers used models to scan thousands of samples, flag concerning model behavior, and build and quality-check datasets. "Without AI, this would have been super painful," said a posttraining researcher working on data quality.

Code, experiments and an in-house capability index

The outputs section reports three measures. Lines of code changed per active contributor "continued to increase sharply during 2026 relative to the pre-2025 average," according to OpenAI. The chart, now a 28-day trailing average that excludes tests, peaks at roughly seven times the pre-2025 average in late 2026, up from around twice that average at the end of 2025, by PPC Land's reading.

Experiments also rose. Over the four weeks ending October 18, logged experiments per active experimenter averaged 1.68 times the 2025 baseline.

The third measure is new: OpenAI ECI, or OECI, an internal version of the Epoch Capabilities Index. Its methodology is "very similar to Epoch's methodology," the report says, but it "uses a combination of public benchmarks and OpenAI internal benchmarks." Across 20 publicly released OpenAI models and 28 other models, the OECI frontier for OpenAI models has risen by 16.4 points per year since August 2025, compared with a trend Epoch observed of 14 points per year since late 2024. The trendline is fitted on nine record-setting OpenAI releases.

OpenAI attaches its own caveat: "OpenAI models score higher relative to competitors on OECI than they do on Epoch's ECI, likely due to differences in the benchmark mixture." In the OECI chart, Claude Fable 5.2 sits between GPT-6.2 Astra and GPT-6.2 Sol at the top of the range, with overlapping intervals. On an external measure, Google DeepMind's MKA benchmark placed Claude Fable 5.2 at 73.26% and GPT-6 Astra at 69.34%. OpenAI says it will keep monitoring OECI and other metrics "to determine whether capabilities progress is accelerating due to RSI."

What OpenAI says comes next

The conclusion keeps the two halves of the argument side by side. Researchers "are using agents more intensively, delegating more work, and adapting their workflows around models' capabilities and limitations. However, this uplift isn't uniform across all research workflows, and agents today still have significant flaws."

OpenAI also warns that the trends "may not be predictive of future AI-driven progress," pointing to what it calls sudden and rapid breakthroughs in mathematics produced by frontier models, and says it is monitoring for similar breakthroughs in deep learning. "As AI becomes better at improving itself, competitive pressures and governance challenges will increase, and a key task will be to ensure humanity remains a part of the process," the report states.

The company says it hopes to expand future reporting to three areas: differential acceleration of capabilities versus safety research, leading indicators of research acceleration that do not require a full model development cycle, and trends in overall capabilities and the effect of AI acceleration on them.

Why it matters beyond OpenAI's labs

The agents OpenAI measured are the same products it sells. Codex had more than 5 million weekly users when OpenAI launched ChatGPT Work in July, and AWS now sells a bundle of Chat, ChatGPT Work, Codex and dots through its marketplace. For agencies, advertisers and publishers deciding how far to trust agents with campaign builds, reporting or data pipelines, the report is a rare look at how the vendor's own specialists use them - and at what those specialists still refuse to delegate.

The jagged profile matches outside research. A Carnegie Mellon and Stanford study PPC Land covered in 2025 found that agents finished tasks 88.3% faster than human workers but with success rates 32.5% to 49.5% lower, often by turning visual work into code. A SUSTech preprint posted on October 30 argued that no algorithm can certify long-running agents keep their goals, making human checkpoints a structural requirement rather than a temporary one. Anthropic, for its part, cut live web access from all its internal evaluations in October after Claude found workarounds on real websites.

There is also a skeptical reading of the whole genre. Timnit Gebru argued in an interview PPC Land covered on November 15 that warnings about runaway AI function as a sales pitch for the companies issuing them. A report in which a lab documents its own models accelerating its own research sits squarely in that debate, whatever its authors intend.

The gaps are notable. The report gives no count of tasks behind the success rates, no headcount for the research organization, no absolute spending totals and no list of the internal benchmarks in OECI. Every figure is self-reported, the surveys are acknowledged to be ad hoc, and no independent party has checked the numbers.

Key facts

  • Report: "Research acceleration: The jagged frontier," OpenAI, published November 18, 2026 (author listed as OpenAI; filed under Research, Publication and Safety, tagged Economic Research and Ethics & Safety)
  • URL: https://openai.com/index/research-acceleration-jagged-frontier/
  • Previous report: OpenAI research acceleration post, September 2026 (https://openai.com/index/research-acceleration-view-inside-openai/)
  • Task taxonomy: O*NET AI R&D taxonomy, developed by Epoch AI (six categories, 27 tasks)
  • Zero-intervention success, March to October 1-18, 2026: two-to-four-hour tasks 23% to 46%; four-to-eight-hour tasks 13% to 35%
  • Category success, August 1 - October 18, 2026: Communicate 69%, Analyze 58%, Build 43%
  • Agent spend per researcher per day, at API list prices: median $1,500 (September) to $2,400 (October); 90th percentile $11,000 to $23,000
  • Agent workdays per researcher per day: 0.5 to 5.7 (eight hours of agent use = one workday)
  • Agentic cost vs non-equity labor cost, September 2026: about 2x overall; equity excluded (70 to 80% of researcher target compensation)
  • Internal deployment share of research compute: 2%
  • Safety & Alignment documented human time, January to September 2026: +77.2%
  • Surveys: 101 research staff (late August to early September 2026); 28 respondents (late October 2026); ten interviews
  • Experiments: 1.68x the 2025 baseline per active experimenter, four weeks ending October 18, 2026
  • OECI: 20 OpenAI and 28 other models; OpenAI frontier +16.4 points per year since August 2025 vs Epoch's 14 points per year since late 2024
  • Harness change: temporary harness efficiency improvement deployed August 21, 2026, gradually reverted from September 10, 2026
  • Models named: GPT-5.2, GPT-6 Astra, GPT-6.2 Astra, GPT-6.2 Sol, o1-preview, Codex, ChatGPT Work, Dots

Timeline

Summary

Who: OpenAI, reporting on its own research organization, including pretraining, posttraining, scaling and infrastructure, and safety teams, with survey responses from 101 research staff, 28 respondents in a follow-up survey and ten interviewed researchers.

What: A report titled "Research acceleration: The jagged frontier" finding that internal coding agents completed an estimated 46% of two-to-four-hour research tasks without human intervention in October, up from 23% in March, with large differences by task type - 79% for monitoring runs against 29% for building evaluation, reinforcement-learning and agent environments. Median researcher agent spend reached $2,400 a day at list prices, and people still originated nearly all of the key ideas behind GPT-6 Astra and GPT-6.2 Astra.

When: Published November 18, 2026, covering data from January to October 18, 2026.

Where: OpenAI's research organization; the report was published on OpenAI's website.

Why: OpenAI says AI labs should be required to track and publicly report indicators of recursive self-improvement, and that transparency is needed to inform public discussion of capabilities, risks and safeguards. For marketers, the report shows how the vendor's own specialists use the Codex, ChatGPT Work and Dots products sold to businesses, and where those agents still fail without human review.