AI Token Costs Now Run $600 a Day for the Median OpenAI Researcher

On September 6, OpenAI published a chart of its own bill.^[1] One line, one axis: daily dollars of inference per researcher, from near zero in February 2026 to about $600 a day by late August. Inference is what it costs to run the model, and the chart prices it per person.

The coverage read the post for its headline claim about agents outworking people. The chart is the more useful artifact, because it is a per-seat price published by the company that collects it.

What did OpenAI actually publish?

A post titled "Research acceleration: The view inside OpenAI," dated September 6, 2026. By mid-August its median researcher was using more than $600 a day of inference at list prices, its 90th percentile user more than $7,000 a day, and its research organization 3.1 agent-workdays of effort for every workday of human labor.

OpenAI also says it has reached the "automated research intern" goal Sam Altman announced last fall. It defines that intern as "a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days," and it is now aiming at a fully automated AI researcher by March 2028.

Its chief scientist published an essay the same day arguing no lab yet knows how to scale safely. This post takes no position on that argument.

What does 3.1 agent-workdays actually measure?

Runtime, not output. OpenAI counts agent hours against a standard eight-hour day, and reports that total agent runtime across the research organization only passed total human labor after June 2026. Agents run overnight and several at a time, so the ratio tracks how much compute is running beside people.

A researcher who leaves four agents running while they sleep produces a large ratio and no finished work until someone reads the results in the morning. The ratio is honest about compute and silent about outcomes.

The spend distribution is the number that describes what an AI-native seat costs. It is the one the coverage repeated and then set down.

How much is $600 a day per researcher?

About $13,000 a month across 22 working days, at the prices anyone else pays for the same tokens. The 90th percentile seat runs past $150,000 a month on the same arithmetic. OpenAI pays its internal cost rather than list, so read both as what the workload costs everyone else.

That caveat matters and it does not rescue the number. List price is the honest benchmark precisely because it is what a customer pays, and OpenAI chose to state the figure that way.

How does that compare with what everyone else spends?

Ramp's AI Index, built from card and bill-pay data across more than 30,000 businesses, put the top 1% of firms at $7,449 per employee per month in June 2026, the top 10% at $611, and the median firm at $11.38.^[3] OpenAI's median research seat is nearly twice the top 1%.

Ramp's economist put it plainly at the time.

TechCrunch asked the question directly when those figures landed in June: are companies actually spending more on AI than on humans? "Not quite yet," it answered, against the roughly $16,000 a month the average software engineer makes.^[4]

Read the ladder from the bottom and the argument is in the geometry. The median firm spends the price of one subscription seat. The top 1% of firms spend less than half an engineer's salary, which is exactly what Ramp said in June.

Then the last two bars. OpenAI's median researcher sits at about four-fifths of that salary, and its 90th percentile is roughly 10 times past it.

The first organization with seats that cost an engineer's salary in tokens is the one selling the tokens. It also has the least reason to economize, so treat it as an upper bound rather than a forecast.

Why is the spread the real budgeting problem?

Because a budget set on the median will be wrong by an order of magnitude on the seats that matter. Inside one research organization the 90th percentile user spends more than 11 times the median. Across firms, Ramp measured a median monthly AI spend of $2,246 against an average of $140,842.^[5]

Averages that far above medians are the signature of a few very heavy accounts. Ramp also found month-to-month swings above 40% to be common even when headcount stayed flat, which is what makes an annual per-head budget useless here.

The spread is the shape of the work. Spend concentrates in the people who can keep four agents productive at once, and doing that well is a skill rather than a setting.

Anthropic's own research puts the ceiling in numbers: developers use AI in roughly 60% of their work, and report being able to "fully delegate" only 0 to 20% of tasks.^[8]

OpenAI reports the same limit from its side: over half of successful four-to-eight-hour tasks needed a person.

Meanwhile the budgets keep rising on faith. Bain's Automation and AI Pathfinder Survey put 951 companies through the question in June: nearly 40% of those that measured cost savings landed in the 0% to 10% band rather than the 11% to 20% they targeted, and 90% of those same companies are raising budgets again.^[6]

How much of the jump is price, and how much is usage?

Some of each, and the split matters before anyone extrapolates the curve. Simon Willison's guess is that the late-July acceleration is internal access to the model later released as GPT-6 Astra, which lists at $10 per million input tokens and $50 per million output, against $4 and $20 for the model before it.^[7]

Link blogAuthor

Research acceleration: The view inside OpenAI

Simon Willison · 6 September 2026

I'm intrigued at what caused that significant acceleration in AI spend per researcher in late July - my best guess is that's when internal employees gained access to the model later released as GPT-6 Astra.

Source: simonwillison.net. Reproduced verbatim; punctuation is the source's. Photo: Paul Downey, CC BY 2.0, via Wikimedia Commons.

Run the arithmetic on that guess. The chart moves from about $160 a day in July to about $600 in late August, a factor of 3.75. A straight swap to the newer model is 2.5 times the price per token on both input and output, which leaves about 1.5 times of genuine extra usage.

So a bill nearly quadrupled while no list price went up. The seat moved to a model that costs more per token and then ran it harder. That is the mechanism we described in July when the price cuts landed: the bill follows what the agent consumes, and the per-token headline is the smaller half of it.

What should a team do with these numbers?

Budget per seat and expect a spread of more than 10 to 1. Put the controls on the heavy seats rather than on the median. Ask for outcomes next to runtime, because an agent-workday says how much compute ran and not how much work finished.

Four moves follow from the distribution:

• Budget per seat, not per headcount. A team of 20 with three heavy agent users is not 20 seats of AI cost. It is three, plus a rounding error.
• Instrument before you cap. You cannot tell a runaway loop from a productive overnight run without knowing which seat spent what, on which task, with what result.
• Ask what finished. Runtime is the easy metric and the least informative. The useful pair is share of long tasks that succeeded and how many needed a person.
• Turn repeated agent work into fixed workflows. The exploration you have already done twice does not need an agent rediscovering it a third time, and structure is itself a token-cost lever.

Then decide which seats are allowed to look like a salary. At these numbers that is a real decision, and the honest version of it is about what a seat actually buys rather than what a plan costs.

What is a usage cap telling you?

That the seat is priced for chat and being used for agents. OpenAI restored the five-hour usage limit on Codex and ChatGPT Work for Plus subscribers on August 24 after lifting it for a few weeks,^[9] and on September 7 users reported the limits tightening again while the $100 Premium seat stayed uncapped.^[10]

One commenter's read was that Premium "is now a more compelling upsell." The cap is the seat's real cost showing through, arriving as a wall instead of a line item.

For three years the question was whether AI would be cheaper than people. OpenAI's chart answers a different one. A person working the new way costs about a person again, in tokens, and most of that bill sits with the few who work it hardest.