OpenAI’s research organization is now running 3.1 eight-hour-equivalent agent workdays for every human workday. That sounds close to saying AI has made OpenAI research three times faster.
The data does not support that interpretation.
The 3.1 figure measures agent runtime in standard eight-hour workdays, rather than discoveries, successful experiments, model improvements, or researcher productivity. The stronger finding is still significant. Researchers can supervise far more machine work in parallel, they are running more experiments, and agents are succeeding on harder tasks more often.
The unanswered question is what happens after all that extra activity enters the research funnel. How much becomes a valid experiment? How much survives evaluation? How much changes the next research decision? How much becomes an improvement worth integrating into a model?
Those are the measurements that would turn agent activity into evidence of research acceleration.
Key takeaways
By mid-August 2026, OpenAI’s research organization was consuming 3.1 agent-workdays for every human workday, measured in standard eight-hour equivalents. That is a runtime ratio rather than a 3.1× research-productivity multiplier.
The median researcher was using more than $600 per day of inference at API prices. The 90th-percentile user exceeded $7,000 per day.
Experiments per active experimenter reached their highest level since OpenAI began tracking the metric in January 2025. OpenAI also says its available compute grew significantly, so Codex adoption cannot claim all the credit.
Agents are moving beyond code generation into troubleshooting, experiment monitoring, analysis, and other research work. High-level planning still accounts for only a small fraction of agent output.
More than half of successful tasks estimated at 4 to 8 human hours still required at least one human intervention during the previous six months.
Teams deploying agents should measure accepted results, human supervision, cost, cycle time, and downstream quality. Agent-hours tell you how busy the machines were, but they do not tell you how much useful work survived.
What OpenAI actually measured
On September 6, 2026, OpenAI published an unusually detailed look at how coding agents are being used inside its research organization.
The company says it has reached the “automated research intern” milestone announced the previous fall. OpenAI defines that intern as a system capable of carrying out well-defined research tasks under human direction, including jobs that would take a skilled researcher a few days. Its next stated target is an automated AI researcher by March 2028.
That is an important distinction. OpenAI is describing an agent that can execute bounded research work under human direction today, while treating a more autonomous researcher as a future target. The current evidence is therefore strongest around delegation and execution, rather than independent research direction.
The internal usage curve has risen quickly.
At the beginning of 2026, median researcher agent use was modest. By mid-August, the median researcher was consuming more than $600 per day of inference at API prices, while the 90th percentile exceeded $7,000 per day.
Before June, total coding-agent runtime was still below total human labor in the research organization. By mid-August, OpenAI calculated 3.1 agent-workdays for each human workday. More researchers were also running four or more agents concurrently, including subagents created by those sessions.
That concurrency explains why the ratio can climb so quickly.
One researcher can start several independent jobs, keep doing human work, then return later to inspect results. An agent can monitor a run while another writes code and a third investigates a failure. The organization can therefore accumulate machine execution hours much faster than it can add researchers.
This is also the direction agent infrastructure is moving more broadly. Long-running workers, persistent state, retries, and orchestration increasingly matter once teams treat agents as durable execution systems rather than isolated chat sessions.
Parallelism is a real capability. It expands the amount of work that can stay in flight. The measurement problem begins when that execution capacity gets translated into a productivity claim.
More on agentic AI
Why 3.1 agent-workdays does not mean 3.1× faster research
Imagine a researcher spends an eight-hour day working while four agents collectively accumulate about 25 hours of runtime.
Those agents might produce five useful experiment implementations. They might produce three useful ones and two dead ends. They might spend much of their time retrying failures, monitoring jobs, debugging infrastructure, or exploring approaches that are eventually discarded.
Every one of those outcomes still contributes runtime.
That is why agent-hours can rise much faster than validated research output. Runtime is an input to the process. It does not tell you what emerged from the other end.
OpenAI acknowledges this measurement problem directly. The company says code volume and related activity metrics are relatively easy to collect but difficult to interpret because their relationship to actual research progress is uncertain. It also warns that overall research progress is unlikely to keep pace with individual activity measures because AI research contains multiple bottlenecks.
An agent-workday therefore answers a useful operational question: How much machine execution can a research organization keep running alongside its people?
The harder productivity question is different: How much faster is the organization discovering improvements that survive evaluation and make its models better?
The distinction matters because research is a funnel. Code has to work. Experiments have to run correctly. Results have to be interpretable. Promising ideas have to survive follow-up testing. A result then has to matter enough to change what the team does next.
The 3.1 ratio sits near the top of that funnel.
The independent analysis linked in the original article reaches a similar interpretation. The ratio is best understood as a capacity multiplier on execution while judgment, direction, and correction remain bottlenecked on humans.
That still matters. A lab able to execute more reasonable experiments per researcher may search a wider space, eliminate weak ideas faster, and spend less human time on routine implementation. Yet the size of the eventual research-speed multiplier depends on how efficiently the organization converts that added execution into useful evidence.
The evidence for acceleration is stronger than the 3.1 headline
Rejecting the 3.1× productivity interpretation does not make OpenAI’s data uninteresting. Several other measurements point toward genuine acceleration.
OpenAI says experiments per active experimenter have increased throughout 2026, reaching their highest recorded level in August since tracking began in January 2025. It reports a correlation with increased Codex adoption while also noting that available compute has grown significantly, which makes the causes harder to separate.
That caveat is important. If researchers have more compute, they can run more experiments even without a comparable improvement in agent capability. Codex adoption and compute growth are moving together, so experiment count alone cannot isolate the contribution from coding agents.
Even so, more experiments per active experimenter is closer to the output side of the pipeline than runtime alone. It suggests that higher agent use is accompanied by more actual research activity, rather than merely longer-running sessions.
Agents are also taking on a broader range of jobs.
OpenAI classified coding-agent usage using Epoch AI’s taxonomy of the frontier AI research workflow. The six-part framework covers deciding what to work on, designing ideas and engineering specifications, building code and datasets, running experiments and infrastructure, analyzing results, and communicating findings.
Every category increased between January and August. Research and infrastructure coding remained substantial, while technical assistance and monitoring runs grew notably. High-level planning remained only a small fraction of agent output tokens.
That mix matters because it shows agents spreading beyond straightforward code generation. Troubleshooting and monitoring can consume large amounts of researcher attention even when they are not the intellectually decisive part of a project. Moving some of that work to agents can free humans to spend more time on interpretation, prioritization, and research judgment.
OpenAI also reports a small organizational signal outside the token counters. Teams that once held office hours to troubleshoot researchers’ experiments saw attendance decline during 2026. One stopped running the sessions entirely. Traffic to a major human technical-support channel also fell, without an identified shift to another human channel. That is consistent with agents absorbing some routine troubleshooting work.
This is closer to measurable productivity because it points to a human bottleneck becoming less demanding.
Still, debugging an experiment faster is only one stage of research. It says little about whether the experiment embodied a strong idea, whether the result is informative, or whether it changes the model-development path.
Humans have become the expensive part of the loop
The most revealing number in OpenAI’s disclosure may sit below the 3.1 ratio.
During the previous six months, more than half of successful tasks estimated to require 4 to 8 hours of human labor still involved at least one human intervention. OpenAI says human steering becomes increasingly important as task complexity rises.
A successful task can therefore remain expensive in human attention.
A researcher supervising six long-running agents does not necessarily receive six finished research tasks at the end of the day. They may receive six work streams requiring review, clarification, debugging, prioritization, or judgment at different moments. The machines can create more opportunities for progress while also creating more points at which someone has to decide what to trust and what to do next.
That changes the economics of deployment.
As machine execution gets cheaper and easier to parallelize, scarce human attention moves toward verification and decision-making. Researchers still choose which questions deserve investigation, which results are interesting, which experiments should continue, and whether a system should be scaled, paused, or deployed.
Popular AI has already seen the same pattern in software development. AI can make generating another pull request almost trivial while leaving maintainers to absorb the verification, correction, and review burden.
Research has the same structural problem, with a different acceptance test. Generating another experiment is valuable when the organization can judge it cheaply enough. If experiment creation scales faster than review capacity, the bottleneck moves rather than disappears.
The key managerial question therefore changes from “How many agent-hours did we buy?” to “How much validated work did each hour of human supervision unlock?”
More on AI in software development:
A better scorecard for AI-agent productivity
Research, development, and engineering teams can borrow the useful part of OpenAI’s measurement while refusing to turn agent-hours into imaginary employees.
The scorecard should follow the funnel from execution to acceptance.
1. Agent runtime tells you available execution capacity. This is where OpenAI’s 3.1 figure belongs. Rising runtime means people can keep more machine work in flight. It is useful for capacity planning, infrastructure demand, and understanding how much parallel execution the organization can sustain. Presented alone, it is a weak productivity measure.
2. Tasks and experiments attempted measure throughput. If one researcher can launch 15 reasonable experiments where five were previously practical, the search space expands. Attempt volume still needs context because cheap experiments can produce cheap dead ends, but it is a stronger operational signal than runtime by itself.
3. Successful tasks measure quality-adjusted throughput. Define success before the run starts. A completed agent session is not automatically a successful task. The task should meet a stated acceptance condition that reflects the work the organization actually cares about.
4. Human interventions measure supervision cost. Count how often someone must step in and how much time those interventions consume. A two-second approval and a 40-minute debugging session should not share one bucket. Intervention frequency without intervention duration can hide the real labor cost.
5. Accepted results measure useful output. In software, this can mean patches accepted after review. In research, it can mean experiments whose results are considered valid, reproducible, and relevant enough to inform the next decision. Popular AI’s coding-agent comparisons use the same logic by focusing on accepted patches, retries, tool calls, token use, latency, and human review time.
6. Compute cost per accepted result measures economic efficiency. OpenAI’s $600-plus daily median shows how quickly agentic work can turn inference into an ordinary research expense. The useful denominator is accepted work rather than tokens or agent sessions. A cheaper run that fails repeatedly can cost more than an expensive run that produces a usable result.
7. Idea-to-decision cycle time measures organizational speed. Track the elapsed time from a research question to enough validated evidence to decide whether the idea should be pursued. This gets closer to the outcome executives actually mean when they ask whether research is moving faster.
8. Idea origin shows whether agents are executing research or beginning to direct it. Track how many accepted improvements began as a human hypothesis, an agent-generated hypothesis, or a joint refinement. OpenAI’s current data still shows high-level planning taking a small share of agent activity, so execution appears further along than autonomous research direction.
This framework also gives teams a more useful cost calculation than advertised token prices. Popular AI’s AI API comparison argues that failed attempts, retries, infrastructure, human review, and repair all belong in cost per accepted task.
The same principle applies to research agents. If a system makes experiment generation almost free but triples the review burden, the headline runtime gain can coexist with disappointing economics. If it increases accepted results while reducing human intervention per result, the case for real acceleration becomes much stronger.
More on AI tokenomics:
Measuring agent productivity gets harder as agents become useful
There is a slightly perverse measurement problem here. Once agents become useful enough to run concurrently, old productivity experiments start breaking.
METR ran a randomized study in early 2025 in which experienced open-source developers took 19% longer on selected tasks when AI tools were allowed. That result applied to the studied developers, repositories, tasks, and early-2025 tools. It was never a universal claim that AI slows software development.
By February 2026, METR believed newer tools were probably providing more acceleration, but its follow-up study had become difficult to interpret. Developers increasingly declined participation because they did not want to work without AI. Some changed which tasks they were willing to submit. Others found task-time reporting difficult because they ran agents while working on unrelated tasks, making the central task-level estimate a poor proxy for real productivity.
That is remarkably close to the measurement problem OpenAI now faces internally.
When a researcher starts several agents, switches to another task, checks one result, redirects a second, and leaves a third running through lunch, “How long did this task take?” stops having an obvious answer.
Wall-clock time for one task can understate the amount of machine execution happening in parallel. Agent runtime can overstate useful output because unsuccessful loops and discarded work still consume time. Human time can also be fragmented across supervision moments that are difficult to assign cleanly to one project.
The unit of analysis therefore has to move upward.
Measure the research pipeline rather than one person sitting at one keyboard. Count accepted results per researcher, time from question to decision, compute per accepted result, and human review minutes per successful task. Those metrics remain imperfect, but they are harder to inflate simply by running more agents for longer.
OpenAI’s data suggests acceleration, but the multiplier is still unknown
The evidence supports a strong but bounded finding: coding agents are expanding the amount of research execution OpenAI can run around each researcher.
OpenAI reports more agent runtime, more concurrent sessions, more experiments per active experimenter, broader delegation, rising task success across several difficulty buckets, and at least one plausible signal that human support work is being displaced by agents.
Calling that “no acceleration” would ignore substantial evidence.
A specific research multiplier remains unsupported.
The missing measurements sit later in the funnel. How many experiments are genuinely informative? How many proposed improvements survive evaluation? How much researcher time is consumed reviewing agent output? How much compute is burned on dead ends? Is the rate of useful new ideas increasing, or is execution mostly getting faster around a similar pool of human-generated directions?
Those questions matter more as OpenAI moves from its current research-intern milestone toward the automated researcher it wants by March 2028.
Today’s system performs well-defined work under human direction. A stronger signal of recursive research acceleration would appear when agents contribute materially to choosing promising research directions, produce hypotheses humans actually pursue, require fewer interventions, and increase the rate of improvements that survive evaluation.
Runtime could remain at 3.1 while research speed rises dramatically if accepted output improves and supervision falls.
Runtime could rise to 10 while the organization runs into a wall of review, compute, evaluation, or human judgment.
That is why the ratio is interesting without being the answer.
The metric to watch after OpenAI’s 3.1 agent workdays
OpenAI has provided strong evidence that AI agents let researchers run much more work in parallel. Its disclosure also gives clear reasons to reject “3.1 agent-workdays” as shorthand for “3.1× faster AI research.”
The next metric should move closer to validated output.
Watch validated research output per human researcher, then account for the human supervision and compute required to produce it. Watch how quickly an idea becomes a reliable decision. Watch the fraction of successful tasks that need intervention. Watch whether agents begin originating research directions that humans actually pursue rather than mainly executing well-defined work.
The decisive signal will be a set of curves moving together: more accepted research output, shorter idea-to-decision cycles, lower supervision per accepted result, and a rising share of useful agent-originated research direction.
If those measurements improve together, AI accelerating AI research stops being an inference from busy machines.
It becomes a measurable feedback loop.
Explore more from Popular AI:
Start here | Local AI | Builds & gear | Autonomy & policy | Fixes & guides | Popular AI podcast