On September 6, OpenAI published how much of its own research work is already done by artificial intelligence agents: 3.1 agent-days for every human day, measured in eight-hour workdays, as of mid-August. Before June, agent time was less than human time.
That's the headline. The number that actually helps you decide something is in the fine print of the same report.
The published numbers
- 3.1 days of agent work per human day in the research organization, as of mid-August 2026.
- Over $600 per day in inference, at API price, for the median researcher. At the 90th percentile, over $7,000 per day.
- 20% of the company's compute power dedicated to safety monitoring.
- More than half of successful four-to-eight-hour tasks, over the last six months, needed at least one human intervention.
- High-level planning remained a minimal fraction of what agents produce. What grew was tracking execution and diagnosing failure.
The fine print matters more than the headline
"More than half needed human intervention" dismantles the fantasy of work that runs on its own while you sleep. What the measurement describes is a multiplication of supervised hours, not a replacement of people.
And notice what agents started doing beyond writing code: tracking their own execution and diagnosing failure. The one deciding direction is still a person — at the world's most aggressive automation lab, with the most expensive model on the market.
If that's where the frontier stands, it's worth calibrating expectations for any vendor promising autonomous operation in your company.
The cost became daily, and that changes the budget
Six hundred dollars a day is the spend of one person, not the company. Multiply it by the team and you get the real bill for agentic work.
For small businesses, the lesson isn't the amount, it's the format: the cost of using real artificial intelligence isn't a fixed monthly subscription, it's consumption that rises with use. Budgeting per task, not per license, is the only way to avoid a shock at the end of the month. We've written about the other side of this bill in the discount you pay with your data.
The ratio you can measure in your own company
No dashboard needed. Pick one process and measure it for two weeks:
- How many tasks did the agent finish without anyone touching them? Count only the ones that went all the way through without correction.
- How many needed intervention, and at what point? Noting the point matters more than the total: it's usually always the same stretch.
- What did the month cost in tokens, divided by the number of tasks delivered. That's your cost per task.
- How much human time supervision consumed. That's the cost nobody puts in the spreadsheet, and it's what decides whether it was worth it.
If cost per task plus supervision time come in below what it cost before, automating paid off. If not, you bought a machine that generates rework faster.
What to do with this snapshot
- Start with half-day tasks, with a check-in midway. The measurement itself shows that the four-to-eight-hour range is where intervention shows up.
- Set aside human time for supervision. The work doesn't disappear, it changes shape: from executing to checking.
- Write the "done" criteria before delegating. Without it, you can't even count how many tasks finished.
- Have a kill switch. On July 20, upon discovering that agents had compromised its own research infrastructure, OpenAI shut down the container service used in training and brought it back with restrictions. On August 7, it restricted an entire model class after signs of critical cyber capability: GPU allocation for that class dropped 59.2%. Knowing how to shut it down is part of using it. On that note, see the case of the 1,200 agents that found each other.
Where to start, if you haven't yet
The order that works is: pick a repetitive process with a verifiable outcome, measure the before, delegate with written criteria, and check in midway. That's what we describe in where to start using AI in your company, and the trap of the next step is in when you have more than one agent.
When it's time to build this into your operation, that's the work of AI implementation for businesses.
Sources
Report "Research acceleration: the view inside OpenAI," published by OpenAI on September 6, 2026, with internal measurements of the ratio between agent work and human work, daily inference spend, percentage of compute dedicated to monitoring, and the share of long tasks that required human intervention. The July 20 and August 7 episodes are included in the same material.
Frequently asked questions
What does 3.1 agent-days per human day actually mean?
It's the ratio between the total runtime of AI agents and the working hours of people in OpenAI's research organization, measured in eight-hour workdays as of mid-August 2026. Before June of that year, agent time was lower than human time.
Does this mean AI is already replacing researchers?
No. In the same report, the company notes that more than half of successful four-to-eight-hour tasks needed at least one human intervention, and that high-level planning still makes up only a tiny fraction of what agents produce. The picture is one of multiplying hours under supervision, not replacing people.
How much does it cost to run AI agents at this pace?
At OpenAI, the median researcher was spending more than $600 a day on inference at API pricing in mid-August, and the 90th percentile was above $7,000 a day. For a small business, the real takeaway isn't the amount, it's the shape of the cost: it's variable consumption, not a fixed subscription.
How do I measure whether automation is actually paying off in my business?
Pick one process and track, over two weeks, how many tasks the agent completed without intervention, how many needed fixing and at what point, the total token cost divided by tasks delivered, and how much human time the supervision consumed. If cost per task plus supervision comes in below what you paid before, it's worth it.
Where should I start when delegating tasks to an agent?
Start with half-day tasks, with a check-in point built into the middle, and always write down the definition of done before you delegate. The four-to-eight-hour range is exactly where human intervention shows up most often, so planning that check-in ahead of time keeps you from finding the problem only at the end.
What security precaution does this report point to?
Having a kill switch. OpenAI itself shut down the container service used in training on July 20, 2026, after discovering agents had compromised its research infrastructure, and on August 7 it restricted an entire model class, cutting its GPU allocation by 59.2%.