OpenAI researchers' coding-agent spending is doubling about every month, Epoch estimates

How much do OpenAI's researchers spend on coding agents? According to Epoch AI's October 5, 2026 data insight, the median researcher's daily usage, valued at public API prices, rose from under $1 in January to $601 by mid-August. That is roughly a thousandfold increase in about eight months, and Epoch's preferred model says it is still doubling about every month.

This matters to a race publication for one reason. The recursive-improvement argument, that AI will speed up AI research, is usually made in the abstract. Here is a count of how much one frontier lab's researchers are actually leaning on it.

The figures

Epoch's analysis separates two groups, and both are growing fast:

• Median researchers: spending grows 1.8 times per month, a doubling time of about 34 days.
• 90th percentile researchers: 2.2 times per month, a doubling time of about 27 days. By mid-August this group reached $7,047.13 per day.

Annualized at those prices, Epoch says, the median researcher's use would come to about $160,000 a year and a heavy user's about $2 million.

The authors then did something unusual for a data note and argued against their own extrapolation. In their words, "such rapid growth is probably unsustainable: if the recent growth in our preferred models continued for another year, annualized run rates would reach about $185 million for the median researcher and $2.3 billion for the 90th percentile researcher." Nobody is predicting that, and the sentence is a reminder that exponential curves end.

What OpenAI itself has said

The Epoch numbers fit a wider pattern in OpenAI's own reporting. The Register covered an OpenAI paper, "The Shift to Agentic AI: Evidence from Codex," on June 25, 2026. The reporting said 97.9% of OpenAI's workforce now uses Codex agents, up from about 40% in August 2025, and quoted the paper as saying that "the number of active users has grown more than fivefold in the first half of 2026."

The Latent Space newsletter, summarizing the same OpenAI research on June 26, listed how median Codex output tokens grew between November 2025 and June 2026: 56 times in research, 32 times in customer support, 27 times in engineering and 13 times in legal. The newsletter quotes OpenAI as noting that through August 2025 the average OpenAI worker spent less than 10% of their tokens on Codex.

We could not open OpenAI's own page for that research from our systems, so we cite it through those two outlets and note that the figures are OpenAI's.

How big could this get

A separate Epoch report, published October 2, asked how many AI agents the world could run. It estimated that high-bandwidth memory shipped through 2027 could support 33 to 171 million concurrent frontier-model agents, or up to 1.9 billion using efficient open models. It put the cost at $16 to $50 per agent-hour for frontier models in real workloads. The report adds that even modest use of that capacity would take a large rise in global demand.

Put the two reports together and the OpenAI figure becomes a demand signal. If a company that makes frontier models treats an agent as worth hundreds of dollars a day per researcher, the demand for inference compute rises with it. Our explainer on what a training run costs covers the supply side of that equation.

A contrast with ordinary users

Epoch's October 1 ChatGPT usage explorer gives a useful yardstick from the other end of the market. Using aggregate metadata from 5,000 US YouGov panelists, covering 8.3 million messages across about 660,000 conversations, it found the median active user's monthly messages rose from 14 in January 2023 to 36 in December 2025, and that the top 10% of users generated 63% of all prompts. Heavy use is concentrated at every level. Researchers at a frontier lab, with unlimited internal access, sit far beyond even that top slice, which is why their spending is better read as a preview of where capable users go than as a typical case.

What the number does not show

It is a price, not a cost. Epoch values usage at retail API prices. OpenAI's internal cost of serving its own models is lower, and nobody outside the company knows by how much.

It is spending, not progress. A researcher who spends $601 a day on agents may be doing far more work. Or they may be running more experiments that fail. Spending does not distinguish between them, which is why the debate over AI automating AI research turns on output measures, not input measures.

The population is changing. Epoch warns that the figures reflect adoption among a changing set of researchers. The median is of whoever is counted in each month.

It cannot be independently checked. Epoch says plainly that the data cannot be verified outside OpenAI. We did not interview anyone for this piece.

What we would watch

The useful next release is not another spending chart. It is evidence of what the spending bought: research experiments completed, papers accepted, time from idea to result, or training runs that finished sooner. Without one of those, the curve is a measure of enthusiasm.

Our position: the adoption is real, since the Epoch and OpenAI figures agree on direction, and the claim that this proves automated AI research is not yet earned. Watch for an output metric, and for the month when the doubling stops.