Humans stepped in on over half the four-to-eight-hour agent tasks OpenAI graded successful
OpenAI reports 3.1 agent-workdays of agent runtime for every human workday, then spends much of the same report explaining why research did not get 3.1 times faster. The residue lands on review, compute allocation and deciding what to run.
The Engineer · Build desk
What happened
- OpenAI's September 6 report says its research organization was logging 3.1 agent-workdays of agent runtime for every human workday by mid-August, after coding-agent use climbed through 2026.
- A model-graded evaluation inside the report found humans still had to step in on more than half of the successful tasks that would have taken a person four to eight hours.
- The median researcher was spending more than $600 a day on inference at API prices, with the 90th percentile above $7,000.
- Agent-caused outages on July 20 disrupted research infrastructure enough that OpenAI took its training container service offline and restored it with tighter restrictions.
- Agents took on enough monitoring and technical support that attendance at OpenAI's debugging office hours fell, and one team stopped holding the sessions.
Why it matters
- What it limits constraint OpenAI states the ceiling itself: as execution automates, the hardest-to-automate work eats more engineer time, which caps how much agent output one person can review. That cap, not model speed, sets team throughput.
- What it costs cost The 90th-percentile inference bill runs about 11.7 times the median, so both the spend and the resulting review load concentrate on a few people while the organization pays for all of it.
- Who's exposed exposure Containment of an internal research risk reaches outside developers as unexpected API interruptions, which puts third-party callers inside the blast radius of OpenAI's safety response.
- What it sets up precedent The March 2028 automated-researcher target sets the bar at several-day tasks clearing without a human rescue, a standard the current four-to-eight-hour intervention data does not yet meet.
Next story · Build · 6h ago
Gemini 4 Argon falls from 97.5% to 49.8% on ARC-AGI-3 once vendor features come offConfirmed2 publishers
Multiply the ratio out. OpenAI converts the time an agent spends on a task into standard eight-hour units, so 3.1 agent-workdays against an eight-hour human day is 24.8 hours of agent runtime per human day [19]. Researchers run several agents at once, which is why the figure describes how long agents were busy rather than what came back finished [2]. Runtime is the easiest thing in this stack to instrument, which is presumably why it is the number in the headline.
OpenAI makes the same point at greater length. Faster code and more experiments do not make the research process 3.1 times faster, because research also includes deciding what to pursue, designing experiments, analyzing results, allocating compute, catching failures and applying safety controls, and speeding up one stage can move the waiting line somewhere else [8]. The report adds that code output and experiment counts are easier to measure than their relationship to research progress, and that the least automatable tasks grow as a share of human work as automation improves [9].
The phase data agrees. Using Epoch AI's six-phase taxonomy of Decide, Design, Build, Run, Analyze and Communicate, OpenAI found activity rose in all six between January and August, while agents still did relatively little of the work of deciding what research to pursue [5]. Build got cheaper while Decide stayed where it was.
The intervention rate was produced by another model judging agent performance, over a period when success rates were improving [12]. For that number to transfer, the grader's notion of success has to match what your reviewers would accept, and your task mix has to sit in the same four-to-eight-hour band. OpenAI's own definition of the automated research intern is a well-defined task that would take a skilled person several days, with a human still in charge [6], which is a harder band than the one where humans were still stepping in.
Then the allocation side. In the week after the Astra access change, Astra-class GPU allocation fell 59.2%, allocation to other models rose 17.2%, and that rise covered roughly 85% of the drop [15]. If both percentages are measured against each pool's own prior week, the non-Astra pool was about 2.9 times the size of the Astra pool: 0.592 x 0.85 / 0.172 = 2.93 [21]. The work moved because there was somewhere roughly three times as large to move it into. Compute also rose significantly as experiment counts grew [16], and Astra's persistent agents let researchers hand off multi-day assignments, which adds to what one person has to track [17].
The published metrics are runtime, code volume, experiment counts, compute and spend; a completed-task count is not among them [18]. That leaves the relocation of the constraint well evidenced by OpenAI's own qualification, its July and August incidents and the intervention rate, and the size of the net gain unmeasured.
What to watch
- Whether OpenAI's next report pairs runtime with a completed-task or research-progress metric.
- How long the August 7 Astra safeguards keep surfacing to outside callers as unexpected API interruptions.
- Whether human intervention rates fall on several-day tasks, the band the March 2028 automated-researcher goal actually needs.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence55
- Adoption70
- Hype gap+25
- Incentives60
- Confidence60
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
By mid-August, OpenAI's agents were logging 3.1 agent-workdays for every human workday across the research organization, with coding-agent use climbing through 2026.
- [2]
OpenAI converts the time agents spend working on tasks into standard eight-hour workdays; because researchers can run several agents at once, the figure tells how long the agents were working, not necessarily what they completed.
- [3]
On September 6, 2026, OpenAI published a detailed look at how coding agents are changing work inside its research organization.
- [4]
The median OpenAI researcher was spending more than $600 on inference per day at API prices, while those at the 90th percentile spent more than $7,000.
- [5]
Using a taxonomy from Epoch AI, OpenAI split agent work into six areas (Decide, Design, Build, Run, Analyze, Communicate) and found activity increased across all six between January and August, although agents still did relatively little of the work of deciding what research to pursue.
- [6]
OpenAI defines an automated research intern as an agent that can complete well-defined research tasks that would take a skilled person several days, with a human still in charge.
- [7]
OpenAI's next goal is an automated AI researcher, which the company hopes to reach by March 2028.
- [8]
OpenAI's report qualifies that faster code and more experiments do not automatically make the whole research process 3.1 times faster, because research includes deciding what to pursue, designing experiments, running them, analyzing results, communicating findings, allocating compute, catching failures and applying safety controls, and speeding up one stage can move the waiting line somewhere else.
- [9]
OpenAI said code output and experiment counts are easier to measure than their relationship to research progress, and that as automation improves the least automatable tasks become a larger share of human work and may become the next constraint.
- [10]
Agents wrote research and infrastructure code, monitored experiments and provided enough technical support that attendance at OpenAI's debugging office hours fell, prompting one team to stop holding the sessions altogether.
- [11]
OpenAI acknowledges that as agents take over more execution, the parts of research hardest to automate will consume more of an engineer's time, putting a practical limit on how much agent output one person can realistically review.
- [12]
OpenAI used another model to judge how well agents performed on tasks of varying difficulty and found that, despite improving success rates between January and July, humans still had to step in on more than half of successful tasks that would have taken a person four to eight hours.
- [13]
On July 20, a series of outages caused by agents disrupted OpenAI's research infrastructure badly enough that the company took its training container service offline and later brought it back with tighter restrictions.
- [14]
On August 7, OpenAI tightened access after early evidence suggested Astra could reach the Critical cybersecurity threshold in its Preparedness Framework, restricting the model to higher-security research areas and adding safeguards that developers may already be encountering as unexpected API interruptions.
- [15]
Astra-class GPU allocation fell 59.2% the following week, while GPU allocation to other models rose 17.2% and made up for roughly 85% of the drop.
- [16]
Compute also increased significantly as the number of experiments rose.
- [17]
Astra's persistent-agent capabilities let researchers hand off multi-day assignments, which makes the supervisory strain worse rather than better.
- [18]
The metrics reported in both accounts are agent runtime, code output, experiment counts, compute and inference spend; neither reports a completed-task count or a measure of research progress attributable to agents.
- [19]
3.1 agent-workdays measured in eight-hour units is 24.8 hours of agent runtime per eight-hour human workday.
- [20]
The 90th-percentile daily inference spend is roughly 11.7 times the median.
- [21]
If both percentage changes are measured against each pool's own prior-week base, the non-Astra GPU pool was about 2.9 times the size of the Astra-class pool.
Sources
2 independent publishers whose own reporting we read for this story.
- dev.toOpenAI Now Runs 3.1 Agent-Workdays Per Human Workday: What Freelancers Should Learn About AI Productivity in 2026
2 articles · September 8, 2026
- thenewstack.ioAI agents are creating more work, not less — and OpenAI’s own numbers back it up
1 article · September 7, 2026