Trader analyzing financial data on multiple monitors in an office setting
LabsBased on company claims

OpenAI says agents now finish 46% of two-to-four-hour research tasks alone, and humans are still in the loop

By Yash Malviya·

Photo: AlphaTradeZone / Pexels

OpenAI's November 18, 2026 report on its own labs says coding agents raised unaided success on two-to-four-hour research tasks from 23% to 46% between March and October. The numbers are self-reported, and the jagged detail matters more than the headline.

OpenAI published a second look inside its own research organization on November 18, 2026, and the headline number is large: agents now complete an estimated 46% of tasks that would take a human two to four hours, without a person stepping in. In March the same measure stood at 23%. For tasks sized at four to eight hours it rose from 13% to 35%.

This is the closest thing the industry has to a public gauge of whether AI is speeding up AI research, which is the core question in any argument about recursive self-improvement. It also comes from the company with the most to gain from a dramatic answer. So the useful reading is the detail, not the headline.

What OpenAI measured

The report, titled "Research acceleration: The jagged frontier," sorts evidence into inputs, process and outputs. On inputs, the median researcher's agent use reached $2,400 a day at API list prices in October, up 62% from $1,500 in September. At the 90th percentile it hit $23,000, up 105% from $11,000.

Time is the cleaner yardstick. OpenAI counts every eight hours of agent runtime as one agent workday, and says the average has climbed from 0.5 per researcher per day to 5.7. Pretraining and posttraining teams averaged roughly eight. By mid-October, more than two-thirds of the researchers it tracked averaged a daily peak of at least four agents running at once, and more than 10% averaged at least sixteen.

As of September, the API-equivalent value of agent use was about twice researchers' non-equity pay. Yet OpenAI says internal deployment, which combines internal Codex, ChatGPT and API use, accounts for only 2% of total research compute. Nearly all research compute goes elsewhere.

Where the agents fall short

The word in the title is "jagged," and the task breakdown shows why. From August 1 to October 18, estimated unaided success was 69% for communicate tasks and 58% for analyze tasks, but 43% for build tasks. At the task level, monitoring training runs came in at 79% and answering technical questions at 72%. Launching training runs scored 35%. Writing code for inference and serving stacks scored 25%, and building evaluation, reinforcement learning and agent environments scored 29%.

“Overall, uplift is jagged, and humans remain a core part of the AI research loop.”

OpenAI, Research acceleration: The jagged frontier, November 18, 2026

OpenAI says the weak spots could become bottlenecks, since overall uplift is limited by the tasks agents handle worst. Its researcher interviews back that up. Agents "rabbit-holed" on unnecessary steps, wrote what staff called "AI slop" pull requests and made invalid analysis assumptions. People still review, correct and redirect.

Close-up of hands coding on a laptop, focusing on programming productivity
Software researchers at work: OpenAI says its staff now run several coding agents at once. Photo: Alicia Christin Gerald / Pexels

Who has the ideas

The most pointed section asks where research ideas come from, because taste may be the hardest skill to automate. OpenAI asked key contributors to name the most important ideas behind GPT-6 Astra and GPT-6.2 Astra. For both releases, posttraining and safety monitoring researchers said people came up with every idea, and models helped implement them. Pretraining researchers credited people for the GPT-6 Astra ideas. For GPT-6.2 Astra they described working in conversation with Codex, and one credited it with identifying key insights.

A survey of 101 research staff, run from late August to early September, asked what they would give up to keep frontier coding agents for three months. The median answers were 40% of team headcount or 30% of team compute. In a separate late October poll, 27 of 28 respondents said they would rather work with the latest internal model than with a mid-career research engineer plus a 2024 model such as o1-preview. Self-reports are a soft measure, and the sample is small, as OpenAI itself notes.

The safety shift, and the policy ask

One finding cuts against the speed narrative. Safety and Alignment saw the largest absolute increase in documented research effort, up 77% in workweeks from January to September, and OpenAI ties the inflection to the Hugging Face incident in July. That tracks with the company's own account in its safety posts, which describe slower releases and paused training.

OpenAI also restates a policy position: "we believe that AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI)." It adds that the reporting "should not be read as an intention to pursue full RSI." In a September 6 post, the lab said it had reached its goal of an automated research intern and was making strong progress toward an automated AI researcher by March 2028. Our explainer on recursive self-improvement covers the strict and loose readings of that target.

“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.”

Researcher 8, pretraining, quoted anonymously in OpenAI's report, November 18, 2026

Outputs: code, experiments and a new index

On outputs, OpenAI tracks lines of code changed per active contributor and experiments run per active experimenter. It says lines of code have continued to increase sharply during 2026 relative to the pre-2025 average, using a 28-day trailing average and excluding tests after a methodology change. It cautions that rising activity does not necessarily mean faster research progress.

To get closer to results, the report introduces OpenAI ECI, an internal version of the Epoch Capabilities Index, as a more direct measure of model capability gains. The methodological changes since September mean the new series are not directly comparable with the earlier post, which is itself a caution about trends.

What to make of it

The data are unaudited, the task estimates are OpenAI's, and researchers may have changed which jobs they hand to agents as the tools improved, which the report concedes can move success rates on its own. Nothing here shows agents generating the winning ideas.

Still, a lab publishing usage, success rates and survey data, including the numbers that look bad for the automation story, is the transparency regime many people have asked for. Our position: the disclosure is worth more than the headline number. The test is whether Anthropic, Google DeepMind and others publish comparable series, and whether an outside evaluator checks the task sizing. Until then, 46% is a claim to track, not a conclusion to repeat.

Frequently asked questions

How much AI research can OpenAI's agents do on their own?

OpenAI estimates agents completed 46% of research tasks sized at two to four human hours without intervention in October 2026, up from 23% in March. For four-to-eight-hour tasks the figure rose from 13% to 35%. These are OpenAI's own internal estimates.

Is OpenAI claiming recursive self-improvement?

No. The report says it should not be read as an intention to pursue full recursive self-improvement, and it describes uplift as jagged with humans central to the research loop. OpenAI does argue labs should be required to track and report RSI indicators.

Which research tasks do agents handle worst?

From August 1 to October 18, 2026, unaided success was 25% for writing inference and serving code and 29% for building evaluation, reinforcement learning and agent environments. It was 79% for monitoring training runs and 72% for answering technical questions.

How much do OpenAI researchers spend on coding agents?

At API list prices, the median researcher used about $2,400 a day in October 2026, up 62% from September. The 90th percentile reached $23,000 a day, up 105%. OpenAI says the average is 5.7 agent workdays per researcher per day.

Who came up with the ideas behind GPT-6 Astra?

Researchers told OpenAI that people originated the ideas they named for GPT-6 Astra, with models helping implement them. For GPT-6.2 Astra, pretraining researchers described developing ideas in conversation with Codex, and one credited it with key insights.

Has anyone independently verified these numbers?

Not in the sources we opened. The figures come from OpenAI's internal measurements, surveys and interviews, and the report says its picture is partial. Independent checks of the task-duration estimates would be the strongest confirmation.

Sources

What each one is, and whose it is.

  1. 1

    Research acceleration: The jagged frontier, OpenAI (November 18, 2026)

    Vendor announcement
  2. Vendor announcement
  3. Vendor announcement