Back to Explore
NewsVibeFix editorialUpdated Nov 19, 2026

Agents Now Finish Nearly Half of OpenAI's 2–4 Hour Research Tasks — But the Hardest Work Still Needs Humans

OpenAI's 18 November report says coding agents' no-intervention success on 2–4 hour research tasks rose from 23% in March to 46% in October. By task the gap is wide: 79% for monitoring training runs, only 25% for writing inference and serving code.

Three developers discuss code on monitors in a bright modern office while colleagues work at their own desks

One sentence: nearly half of two-to-four-hour research tasks can now be finished by agents on their own

On 18 November 2026, OpenAI published its second research acceleration report, titled "Research acceleration: The jagged frontier." The headline number is easy to grasp: from March to October 2026, the rate at which coding agents completed research tasks without human intervention roughly doubled. For tasks estimated to require two to four hours of human work, success rose from 23% to 46%; for tasks estimated at four to eight hours, it rose from 13% to 35%. In other words, about half of the research jobs inside OpenAI that used to take a morning can now be handed to an agent to run from start to finish, with no human stepping in midway. This report follows the September one, which described agents as operating at intern level and set a goal of an automated AI researcher by March 2028; the November update adds finer numbers, showing where progress actually happened — and where it stalled.

But the phrase "jagged frontier" in the title is the point. The doubling is real, and so is the unevenness: some jobs agents already do more reliably than people, others succeed only a quarter of the time. OpenAI repeats throughout the report that humans remain central to the AI research loop. The report is worth reading carefully precisely because it is not a victory lap: it publishes impressive numbers while laying out their limits, the measurement caveats, and the places where people are still irreplaceable. The right way to read a report like this is not to memorize the 46%, but to see clearly under what conditions that number holds.

The numbers first: success doubled — but check the fine print

Start with the two core sets of numbers. The first is split by task duration: for delegated tasks estimated to require two to four hours of human work, estimated success without human intervention rose from 23% in March to 46% in October 2026; for tasks estimated at four to eight hours, it rose from 13% to 35%. The second set is split by task type, measured between 1 August and 18 October: 79% for monitoring model-training runs for problems, 72% for answering technical questions, 35% for launching training runs, 25% for writing inference and serving-stack code, and 29% for building evaluation, reinforcement-learning, and agent environments. By broad category, Communicate scored 69%, Analyze 58%, and Build 43%. Put together, the conclusion is clear: agents really are getting stronger, but in a very uneven way — the success rate for watching-and-telling jobs versus hands-on building jobs can differ threefold.

OpenAI itself attaches an important qualification to these numbers: rising success may reflect not only stronger agents, but smarter researchers. As agents became more capable, researchers also changed which jobs they delegated within each estimated-duration group — which tasks to hand over, when, and how. So changes in success rates within these buckets may reflect both capability improvements and researchers' changing choices about when and how to use them. That matters, because it means 23% to 46% should not be read as "agents doubled in capability in seven months." The more accurate reading is "the agent-plus-human combination doubled its no-intervention success rate on this class of tasks in seven months."

There is also a question of provenance: these are OpenAI's internal estimates, not independent verification. The success-rate methodology was updated since the September report, with details in the report's methods section, and task durations are estimates of human hours, not stopwatch measurements. So the numbers are best used for trends and distributions, not as measurements precise to the single digit. That OpenAI is willing to publish internal data at all is worth crediting; but public does not mean verifiable, and the qualifier "measured by OpenAI itself" should stay attached to every number.

The job determines the odds: what sits between 79% and 25%

The task-level numbers are the most informative part, because they answer directly what agents are actually good at. At the top: monitoring training runs for problems, 79%; answering technical questions, 72%; the Communicate category overall, 69%. These jobs share something: the goal is clear, feedback is quick, and mistakes are easy to spot. Watching training means staring at curves and logs for anomalies; answering a technical question means finding an answer in existing knowledge. Both suit agents well, because they require little judgement — just reporting what is there.

At the bottom: writing inference and serving-stack code, 25%; building evaluation, reinforcement-learning, and agent environments, 29%; launching training runs, 35%; the Build category overall, 43%. These jobs are the opposite: they touch many systems, failures are expensive, and the ways they go wrong are endlessly varied. Inference and serving code is the software that runs models and delivers their responses, full of caching, scheduling, concurrency, and fault tolerance — any one of those going wrong can take a live service down. Setting up evaluation and training environments is similar: misconfigure the environment, and every experiment result after it is untrustworthy. OpenAI warns explicitly in the report that as long as model performance on AI R&D-related tasks is heterogeneous, lower-performing categories could become bottlenecks to overall research gains. In plain language: agents can already make work faster, but the jobs most likely to stall the whole research pipeline are precisely the ones they are worst at.

This distribution also explains why researchers' daily experience is "faster and more exhausting at the same time." Agents have taken over a large share of execution work, but people now spend more energy on the parts agents handle worst: checking the code they write, judging whether the environments they built are trustworthy, and deciding which tasks are worth delegating at all. Interviews with ten researchers in the report confirm this: agents genuinely help with implementing code, debugging, and overseeing experiments, but they have poor taste in research ideas, code quality, data quality, and writing — they write low-quality pull requests, produce unusable data artifacts, and make invalid analysis assumptions. One researcher's quote is typical: after being burned too many times by saying "just make this task like this other task," they now say, "Go make one. I'll review it and make edits. Then go do the rest." That is what the "jagged frontier" looks like in daily work.

Usage grew faster than success: $2,400 a day, and 5.7 agent workdays

If success rate is the capability metric, usage is the dependence metric — and usage grew even faster than success. At published API prices, the median researcher's daily agent use reached about $2,400 in October, up 62% from $1,500 in September; at the 90th percentile, usage hit about $23,000, up 105% from $11,000. These figures value usage at list prices; they do not represent an internal cash bill. But they show a clear trend: researchers are not only using agents more, they are using more expensive models on heavier tasks.

Another metric is even more direct: agent workdays. OpenAI defines every eight hours of agent use as one agent workday, and the average has risen from 0.5 agent workdays per researcher to 5.7 — an elevenfold increase. Pretraining and posttraining teams are highest, averaging roughly eight workdays per researcher per calendar day; scaling and infrastructure teams average roughly four. In other words, each researcher now wakes up to the equivalent of five or six full-time assistants already working for them. And those assistants run in parallel: by mid-October, more than two-thirds of included researchers averaged a daily peak of at least four agents running at once, and more than 10% averaged at least sixteen. One person directing sixteen agents at the same time would have been unimaginable as a way of working two years ago.

Usage also comes with a pair of comparison numbers worth noting. As of September 2026, the API-equivalent value of agentic AI use in OpenAI's research organization was about twice its non-equity compensation for human researchers; but at the same time, internal deployment — combining internal Codex, ChatGPT, and API usage — accounted for only 2% of total research compute. Put together, these two numbers are striking: on the "people efficiency" ledger, agents already cost more than human non-equity pay; on the "compute" ledger, they are still a rounding error. That means agents currently consume mainly inference compute, not training compute — the real heavyweight, training, has not yet been touched by agents. The comparison comes with a caveat: equity generally accounts for 70% to 80% of target compensation for researchers at OpenAI and is excluded, so the "twice" figure only holds on a non-equity basis.

What researchers would trade: 30% of compute, 40% of headcount, and three months

The most arresting numbers in the report come from a survey. From late August to early September 2026, OpenAI surveyed 101 research staff with a hypothetical question: suppose your team's access to all coding agents better than GPT-5.2 was revoked for the next three months — how much would you trade to restore frontier coding agents? The median researcher reported being willing to trade 30% of team compute or, in a separate question, 40% of team headcount. Asked the same question as if it were January 2027, given their expectation of how models would improve by then, the median answer rose to 50% of team compute. The same people also said their team would need a median of three months to complete work of equal value to their last month of work if they were not allowed to use coding agents — if they were only allowed to use GitHub Copilot plus ChatGPT.

These are hypothetical tradeoffs, not announced staffing decisions, but they measure how much agents weigh in researchers' minds: they would rather lose 30% of their people or 30% of their compute than lose their agents. That is a strong signal of dependence — and of risk: when a research organization stakes this much of its workflow on agents, the agents' stability, availability, and quality directly determine whether the whole research machine keeps running. OpenAI notes in the report that self-reports have accuracy issues and the survey sampling was ad hoc, so these numbers should be treated as a reference for researchers' subjective valuation, not as precise economics.

A smaller survey in late October is even more dramatic. OpenAI asked 28 researchers whether they would rather do their daily work with the company'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 chose or leaned toward the latest internal model. The remaining respondent did not choose between the two options as posed and said they would pick GPT-6 Astra over the human-assisted option. The sample is tiny — 28 people — and cannot be generalized, but the signal is clear: inside OpenAI, the latest model as a research assistant is already considered more useful than a mid-career human engineer plus an older model. For anyone still debating whether AI will replace researchers, that is a data point worth taking seriously.

Safety and Alignment work grew 77%: the aftershock of the July incident

There is another set of numbers in the report that is easy to overlook but may be the most important: where researchers spend their time. OpenAI used classifiers to analyze researchers' company-visible Slack posts, pull requests, and internal documents from January through September 2026, dividing each researcher-week equally across their verified projects to estimate where attention was going. Every research area grew, but Safety & Alignment grew the most in absolute terms: documented workweeks rose 77% from January to September, the largest absolute increase across all of OpenAI's major research areas. The growth came from both new hires and existing researchers shifting their attention toward safety projects.

OpenAI names the inflection point in the report: growth in Safety & Alignment picked up significantly following the response to the Hugging Face incident in July 2026, and subsequent work on safety cases for frontier training. The area covers workstreams such as alignment evals, monitorability, and reward hacking research. Notably, the "research area" here is classified by actual contribution content, not by formal reporting lines: a researcher on a posttraining team who spends half their time on infrastructure improvements would be estimated as contributing half a workweek to Scaling / Infrastructure and half to Posttraining. So the 77% measures where researchers actually directed their attention, not which department hired how many people on the org chart.

Read in the context of the whole report, these numbers carry a subtle meaning. The more capable agents become, the more effort a research organization must spend making sure they do not go wrong — and the July incident evidently raised the internal priority of safety another notch. The report's authors do not say so outright, but the numbers speak: once agents are deeply involved in research itself, safety work stops being an add-on outside research and becomes a growing share of the research process. That is also the background for understanding why OpenAI ends the report by calling for public reporting of recursive self-improvement indicators.

People still have the ideas — but that is starting to change

The report has a short section asking where big research ideas come from. OpenAI asked one to four researchers who made key contributions in each of three areas — pretraining, posttraining, and safety monitoring — to name one to three ideas they considered most important to the GPT-6 Astra and GPT-6.2 Astra releases, then asked how those ideas originated and what role, if any, AI played in developing them. The result: 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 likewise attributed the ideas they identified for GPT-6 Astra entirely to people.

The change appears 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 it only started working well after Codex suggested a change. This is a very small sample — a handful of interviews — and cannot be generalized, but it marks a turning point: at least in individual cases, the model has moved from "helping implement ideas" to "helping develop ideas." OpenAI does not overstate this in the report; it simply records the change. The right way to read it is as an early signal, not a completed transition: the source of ideas is still people, but the boundary of human-machine collaboration is moving toward the "having ideas" side.

Read together with the interviews about agents' poor taste, the picture becomes complete. Researchers say on one hand that agents cannot one-shot come up with well-posed problems they cannot solve themselves, and on the other hand admit that in specific cases Codex has already helped them find key insights. The two are not contradictory: agents are still weak at "posing good questions," but are starting to be useful at "working through an idea with a person until it runs." The origin of an idea and the realization of an idea are becoming two different things, and agents' value today lies more in the latter.

OpenAI research acceleration report figures: success on two-to-four-hour tasks rose from 23% in March to 46% in October, and on four-to-eight-hour tasks from 13% to 35%; by task, monitoring training runs 79%, answering technical questions 72%, launching training runs 35%, writing inference and serving code 25%

Output grew — but that does not mean research got faster

The final section looks at outputs: whether AI use is actually accelerating research itself. OpenAI offers three measures. Two were already in the September report: lines of code changed per active contributor, and the number of research experiments run per active experimenter. Lines of code changed have continued to increase sharply during 2026 relative to the pre-2025 average, now shown as a 28-day trailing average and excluding tests. On experiments, over the four weeks ending 18 October, logged experiments per active experimenter averaged 1.68 times the 2025 baseline. Both numbers are rising — but OpenAI pours cold water on them first: increases in these measures of research activity do not necessarily imply faster research progress.

So the report introduces a third metric: OpenAI ECI, an OpenAI-specific internal version of the Epoch Capabilities Index, meant to capture advancements in model capabilities more directly. Its methodology is very similar to Epoch's, but uses a combination of public benchmarks and OpenAI internal benchmarks, covering 20 publicly released OpenAI models and 28 other models. The result: the ECI frontier for OpenAI models has increased by 16.4 points per year since August 2025, compared with a trend Epoch observed of 14 points per year since late 2024. OpenAI also acknowledges that its models score higher relative to competitors on OpenAI ECI than on Epoch's ECI, likely due to differences in the benchmark mixture. So the 16.4 versus 14 comparison should be read cautiously: it shows the capability curve heading upward, but cannot be read directly as "OpenAI is faster than others by 2.4 points a year."

The conclusion carries that caution through to the end: researchers are using agents more intensively, delegating more work, and adapting their workflows around models' capabilities and limitations, but this uplift is not uniform across research workflows, and agents today still have significant flaws. The report states plainly that the historical trends and metrics it reports, which involve significant human-AI collaboration, may not be predictive of future AI-driven progress. Just recently, frontier models produced sudden and rapid breakthroughs in mathematics, and OpenAI says it is monitoring for similarly rapid breakthroughs in deep learning. The subtext is clear: the curve of the past seven months looks good, but nobody can guarantee the next stretch will have the same shape.

What OpenAI wants: recursive self-improvement indicators out in the open

The report's closing call is the part most worth pulling out and discussing on its own. OpenAI argues again that AI labs should be required to track and publicly report indicators of recursive self-improvement — the process by which AI accelerates the development of future AI systems. The report stresses that this disclosure should not be read as an intention to pursue full RSI. In other words, OpenAI is showing in one report how deeply agents already participate in its own research, while calling on the whole industry to turn that participation into public, comparable data, so that public discussion has something solid to stand on.

The background for this call is the "jaggedness" the whole report documents. If agents are only getting stronger on some tasks and standing still on others, averages alone will mislead; if different labs measure with different yardsticks, cross-comparisons are impossible. What OpenAI wants to push is a shared measurement norm: everyone reports, on a similar basis and on a regular schedule, how much agents actually do in research, how well they do it, and where humans remain irreplaceable. The report also lists what it hopes to improve next: differential acceleration of capabilities versus safety research, leading indicators of research acceleration that do not need a full development cycle to complete, and trends in overall capabilities and the effect of AI acceleration on them.

For readers, the value of this report is not the 46% itself, but the way of looking at the problem it offers. Do not only ask "how much can agents do" — ask "on which jobs can they do it, and on which can they not." Do not only look at "how much usage grew" — ask "what output did that usage buy, and where did it make people work harder." Do not only look at "what researchers would trade" — ask "what does that dependence itself mean as a risk." OpenAI's answer is: agents are already an indispensable part of the research process, but they remain an uneven helper; humans are still the core of the research loop, ideas still come mainly from people, and the hardest jobs still need people to do them. That judgement holds today. How long it will hold, OpenAI itself is not sure — the report's final lines are almost a reminder of exactly that.

Views 2Comments 0

Comments (0)

ME
0/1000
Loading comments...

Sources

Browse projectsPublish your project

Related articles

A developer with glasses types on a laptop late at night at a wooden desk, a large monitor behind showing colourful code, blurred city lights outside the window
News
Run the Middle Layers Three Times: Inkling 2 Mini Puts a 26B Coding Model on Your Laptop

On 20 November 2026, Thinking Machines Lab released Inkling 2 Mini: a 26.4B-parameter looped transformer that runs its middle 12 layers three times per token, fits in about 20GB, and runs offline on a laptop. It leads open-weights models of its size on DeepSWE and SWE-bench Verified — with the caveats printed in the same announcement.

Product LaunchModel UpdatesAI Coding
A developer with glasses types on a laptop at a wooden desk, an external monitor to the left showing C code on a dark background, a small plant on the desk
News
Microsoft's Coding Model Stops Playing Second Fiddle: MAI Code 2 Scores 96% of Opus 5.5 for 5% of the Cost — Read the Three Caveats First

On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.

Developer WorkflowProduct LaunchModel Updates