Technology•

OpenAI's Heaviest Agent Users Now Run Up $23,000 a Day in Tokens, and Staff Would Give Up 40% of Headcount to Keep the Tools

An OpenAI report dated November 18 puts daily coding-agent usage among its heaviest research users at $23,000 in tokens at list prices, up 105% from September.

Written By :MINRK
OpenAI's Heaviest Agent Users Now Run Up $23,000 a Day in Tokens, and Staff Would Give Up 40% of Headcount to Keep the Tools

Photo by Fotis Fotopoulos on Unsplash

1. OpenAI coding agents now cost $23,000 a day at the top end

OpenAI coding agents are consuming about $23,000 worth of tokens per day for the company's heaviest research users, a 105% increase from September, according to an OpenAI report dated November 18, 2026. The document, titled "Research acceleration: The jagged frontier," values that usage at API list prices and tracks how deeply agents are embedded in the work of its research staff.

The report offers an unusually detailed look at how one of the leading developers of artificial intelligence uses its own products. Coding agents are AI systems that can write, run and revise software with limited supervision, carrying out multi-step tasks on a person's behalf. OpenAI sells access to such systems to outside customers, and it also relies on them internally. The new figures describe the second of those uses: how much its own researchers lean on agents to do research.

The answer, by the company's own measurements, is a great deal, and considerably more than two months earlier. Usage roughly doubled among the most intensive users over that period and rose sharply for the typical researcher as well.

2. How much do OpenAI researchers spend on tokens?

At the 90th percentile of OpenAI's researchers, daily agent usage reached $23,000 in tokens at API list prices, while the median researcher's usage reached $2,400 per day. The first figure is up 105% from September and the second is up 62%.

The choice of the 90th percentile is worth noting. It describes the heaviest users of coding agents on the research staff, but not the extreme outliers. A researcher at that level uses more than nine in ten colleagues, yet there remains a further tenth of the staff above that mark. The figure is therefore a measure of what intensive use looks like inside the company, not a record set by a single individual.

The median tells a different part of the story. It marks the midpoint of the staff, the level that half of researchers exceed and half fall below. At $2,400 per day, the typical researcher is consuming a quantity of tokens that would represent a substantial bill for an outside developer, even though it is roughly a tenth of what the heaviest users go through.

A token is the unit in which AI models process and generate text and code. Model providers charge customers by the number of tokens sent to a model and returned by it, which is why usage can be expressed in dollars.

3. Why the list-price qualifier matters

Every dollar amount in the report carries the same condition: it is stated at API list prices. That means the figures reflect what an outside customer paying OpenAI's public rates would be charged for the same volume of usage.

It is a way of translating internal consumption into a unit that people outside the company can interpret. It does not describe what the usage costs OpenAI itself. A company running its own models on its own infrastructure does not pay itself the published rate, so the $23,000 figure should be read as a measure of volume, expressed in the currency of the public price list, and not as a line item in an expense report.

The qualifier also explains why the numbers are useful for comparison. An enterprise weighing whether to give its own engineers the same level of agent access can look at the list-price figure and see approximately what that pattern of use would cost.

4. Researchers are running several agent workdays in a single day

Spending is only one way the report measures adoption. Another is parallelism, meaning the number of agents a single person keeps working at once. The report puts the average at 5.7 agent workdays per researcher per calendar day.

In practical terms, the average researcher is directing the equivalent of more than five full working days of agent activity every day. That is possible because agents, unlike people, can be run side by side. A researcher can hand one agent an experiment to set up, another a piece of code to rewrite and a third an analysis to run, then review the results as they come back.

The report also describes what those agents achieve. Agents now complete almost half of the research tasks that would typically take a human two to four hours. Tasks of that length are not trivial requests. They are substantial pieces of work, the kind that would occupy a skilled person for a good part of a morning or an afternoon, and the report indicates that agents finish them close to half the time.

Taken together, the two measures describe a change in how research time is spent. A researcher who oversees several agents at once is doing less of the hands-on work personally and more of the directing and checking: deciding what should be attempted, dividing it into pieces an agent can handle, and judging whether what comes back is correct. The completion figure cuts both ways, though. If agents finish almost half of the two-to-four-hour tasks, they also fail to finish a little more than half, which means human review and intervention remain a regular part of the process even for the staff who use the tools most.

5. Staff would trade compute or colleagues to keep agent access

One finding speaks to how researchers themselves value the tools. According to the report, staff would give up 30% of their team's compute, or 40% of their team's headcount, in order to keep access to frontier agents for three months.

Both sides of that trade are significant inside an AI laboratory. Compute, the processing power used to train and run models, is the scarcest resource in the field and the one that research teams compete hardest for. Headcount is the other traditional measure of a team's capacity. A stated willingness to surrender nearly a third of the first or two-fifths of the second suggests that researchers regard agent access as more valuable than a large share of either.

The finding should not be over-read. It does not mean OpenAI plans layoffs. It records a hypothetical preference, an answer to a question about what staff would choose if forced to, and nothing in the report indicates a change in hiring. What it does provide is a quantified statement of how the people closest to the technology weigh it against more conventional resources.

6. The gap between heavy and typical users is widening

A comparison of the two growth rates points to an uneven pattern of adoption. Usage at the 90th percentile climbed 105% from September, far faster than the 62% rise at the median.

When the top of a distribution grows faster than its middle, the distance between the two increases. In this case it means the most intensive users are pulling further ahead of their colleagues instead of the rest of the staff catching up with them. Both groups are using agents more than they did, but the researchers who were already running the most are expanding their use at a quicker rate.

That divergence fits the report's title. A jagged frontier is one that advances unevenly, further in some places than in others. The adoption data shows the same shape among people: a group of researchers whose use of agents is accelerating rapidly, alongside a majority whose use is growing strongly but more slowly.

7. What the figures mean for the wider AI industry

The most direct implication is that OpenAI is using its own agents to speed up its own research, and that the internal usage curve is steep. A company that builds AI systems is applying those systems to the task of building the next ones, at a rate that has roughly doubled for its heaviest users since September.

For people watching the effect of AI on employment, the headcount trade-off deserves attention. It reframes the usual question about staffing and tooling. The conventional view treats software as something that supports a team of a given size. The finding treats agent access as something researchers would, hypothetically, rank above a large portion of the team itself. Again, the report describes a stated preference and not a plan.

There are also reasons for caution in applying the numbers elsewhere. OpenAI's researchers are not typical users. They have privileged access to new models, deep expertise in prompting them and no meaningful cost barrier to heavy use. An average enterprise developer, whose employer pays list price for every token, faces a very different calculation. A level of usage that makes sense for a researcher who bears no marginal cost may be hard to justify for a business that does.

8. What to watch next

Whether outside organizations follow the same path will depend heavily on how token prices move. At list prices, the usage described for the 90th percentile of OpenAI's researchers would amount to $23,000 per person per day, and even the median would come to $2,400. Few companies could extend that to an entire engineering organization at current rates.

If prices fall, the kind of heavy, parallel agent use that OpenAI's staff now treat as routine becomes more accessible to others. If they do not, the practice may remain concentrated among the model developers themselves and the best-funded customers. Either way, the report establishes a reference point: a measure of how much agent capacity the researchers with the fewest constraints choose to use, and how quickly that amount is growing.

Related Articles

NEWSLETTERS

Don't miss another story.

Subscribe to the MINRK Newsletter today.

Tick the newsletters you want. No selection? You'll get MINRK Headlines.

By signing up, you will receive emails about MINRK products and you agree to our terms of use and privacy policy.

MINRK Daybook

Market analysis for crypto traders and investors.

EVERY WEEKDAY

MINRK for Advisors

Defining crypto, digital assets and the future of finance for financial advisors.

EVERY THURSDAY

MINRK Protocol

Exploring the tech behind crypto one block at a time.

WEEKLY

MINRK Long & Short

A must read for institutions. Insights, news and analysis delivered weekly.

EVERY WEDNESDAY

MINRK Headlines

The biggest crypto news and ideas of the day.

EVERY WEEKDAY

MINRK Policy

Examining the intersection of cryptocurrency and government.

WEEKLY

MINRK Research

Join thousands of readers who rely on MINRK for data-driven insights on the latest digital asset trends.

MONTHLY