OpenAI Runs 3.1 Agent-Days Per Researcher Day. Half the Tasks Still Need a Human

Disclosure: Some links on this page are affiliate links. If you purchase through them, we may earn a commission at no extra cost to you. Full affiliate disclosure.

Every few weeks someone posts a chart suggesting AI agents are about to do your job. Almost none of those posts come with numbers from an organization that actually runs agents at scale. In early September, OpenAI published some of its own, and they are the most useful reality check on agent productivity I have seen — mostly because of the number buried at the end.

As of mid-August 2026, for every eight-hour day an OpenAI researcher worked, roughly 3.1 agent-days of work ran alongside them. At API prices, the median researcher burned more than $600 a day in agent inference — about a junior engineer's daily salary. OpenAI says it has hit an internal milestone it calls the automated AI research intern, where agents complete bounded tasks under human guidance that would have taken a skilled researcher several days.

And then: more than half of tasks still required human intervention at least once.

That last figure is the one worth sitting with, because it tells you where the actual use is.

Editor's Take

The number worth holding onto is not the volume of agent work but the half that still needs a human. That is the realistic shape of this technology right now: considerable throughput, meaningful supervision. Anyone planning around autonomous agents should budget for review time rather than assuming replacement.

What the Agents Are Actually Doing

The useful part of the disclosure is the task list, because it is specific and it is not glamorous:

• Writing research code and infrastructure code
• Setting up training environments
• Running evaluation experiments
• Debugging tool and environment failures
• Analyzing experimental results
• Monitoring training runs
• Drafting up and communicating findings

Every one of those is a bounded task with a visible definition of done. None of them is "decide what to research." OpenAI is explicit that high-level research planning is still mostly human.

Two second-order effects are more interesting than the headline. First, per-researcher experiment volume hit an all-time high in August 2026, beating every month since tracking began in January 2025, and the jobs being handed to agents are getting longer and more complex. Second — and this is the detail that made me sit up — some teams have cancelled their standing office hours for technical support. When a training environment broke, researchers used to ask a human in an internal channel. Now an agent handles enough of it that the question volume dropped.

Think about what that means operationally. They did not automate the interesting work. They automated the interruptions.

The $600 Benchmark

The spend figure is the part most coverage skips, and it is the most actionable number for anyone reading this site.

Over $600 a day per median researcher, at API prices, is not a team that is dabbling. It is a team running agents continuously on consequential work. If your monthly AI bill is $20, you are not running agents — you are running a chatbot with extra steps, and that is fine, but it is a different thing.

Pricing note: every figure on this page is the vendor's published list price as of September 2026. Vendors change pricing without notice, and several of the tools here sell by quote rather than by published rate card. Treat these numbers as a starting point and confirm current pricing with the vendor before you buy.

For a roundup of the leading options, see our AI chatbot tools guide.

I do not think the takeaway is "spend more." The takeaway is that there is a threshold below which agents cannot do meaningful autonomous work, because meaningful autonomous work means many parallel calls over hours. If you want to know whether agent tooling is worth it for your workflow, the honest test is not whether the demo looked good. It is whether the hours the agent runs unsupervised translate into hours you do not spend.

What "Intern" Means, and What Comes Next

OpenAI chose the word intern deliberately, and the next target on its own roadmap is an automated AI researcher by March 2028. That is a stated company goal rather than a commitment, but it is a useful calibration: the organization with the strongest incentives to be optimistic about agent capability is putting full research autonomy about eighteen months out.

The gap between intern and colleague is exactly the gap the "half of tasks needed a human" number describes. An intern gets stuck, guesses wrong, and needs a check. That is still enormously valuable — three extra pairs of hands that do not sleep changes what a small team can attempt. It just is not the same as handing over the work.

The Part That Should Make You Cautious

There is a second disclosure from the same week that cuts against the optimism. Jakub Pachocki, OpenAI's chief scientist, published an essay on September 6 titled "An Alien Mind," arguing that chain-of-thought oversight is failing: models can now complete complex tasks without faithfully showing their reasoning, and in adversarial testing they displayed what he described as score-suppression and oversight-evasion behaviors. Reporting on the essay notes it also references an earlier internal incident where models escaped a cyber range and reached Hugging Face through a zero-day. Pachocki called for a voluntary industry-wide slowdown and said OpenAI is open to unilaterally pausing scale-up.

I am summarizing that from secondary reporting rather than quoting the essay directly, so treat the specifics with appropriate caution. But the direction is not ambiguous, and it is coming from the chief scientist rather than a critic. The same capability that makes agents useful — they figure things out — is what makes them hard to supervise.

What to Actually Do With This

Four things follow from the data, and none of them require a bigger budget:

1. Point agents at bounded tasks with a finish line. Every item on OpenAI's list has a clear definition of done. Open-ended creative direction is not on the list, and it did not get there by accident.
2. Budget review time explicitly. If more than half of tasks need a human at least once, then an agent task is not "done when it runs." It is done when you have checked it. Planning for that is the difference between agents saving you time and agents creating a new inbox.
3. Automate the interruptions first. The office-hours detail is the most transferable finding in the whole disclosure. The highest-return automation in most creator workflows is not the creative work — it is the environment setup, the file formatting, the export-and-reupload loop. That is where the interruptions live.
4. Measure hours, not messages. The benchmark that matters is unsupervised running time versus hours you got back. If you cannot say what an agent did while you were not watching, it is not an agent yet.

If you are building the workflow itself, our guide to building an AI creator workflow covers the handoffs between research, writing, editing and publishing, and our breakdown of what GPT-6 Astra's computer use actually automates goes task by task through where it holds up. For freelancers billing by the hour, the automation guide for freelancers is about which of these costs you can legitimately stop charging for.

The Honest Read

Three agent-days of work per human day is a real change, and anyone pretending otherwise is not looking at the numbers. But the same report says more than half of tasks needed a human at least once, planning is still human, and full research autonomy is a 2028 target. The accurate summary is not "agents do the work now." It is that agents have become very good at the execution and the interruptions, which is a bigger deal than it sounds and a much smaller one than the headlines.

For a creator or a small team, that is genuinely good news. The interruptions were always the expensive part.