I read OpenAI's monthly-doubling coding-agent curve as demand, not as a cost line

Epoch AI finds OpenAI researchers doubling coding-agent use roughly every month, and the list-price dollars attached to that curve measure usage, not what OpenAI pays.

AI InfrastructureThe Current
Editorial image for I read OpenAI's monthly-doubling coding-agent curve as demand, not as a cost line

Key argument

Epoch AI finds OpenAI researchers doubling coding-agent use roughly every month, and the list-price dollars attached to that curve measure usage, not what OpenAI pays.

$601 a day. That is the coding-agent inference the median OpenAI researcher consumed by mid-August 2026, valued at OpenAI's API list prices, according to Epoch AI's reading of two charts in a September 2026 OpenAI post. In January 2026 the same median figure sat under $1, Epoch AI reports. The 90th-percentile researcher was above $7,000 per day by mid-August on the same accounting.

Epoch AI's headline is that coding-agent use at OpenAI is doubling roughly every month. The sentence I keep returning to is a different one. Epoch says plainly that these figures price inference at list rather than internal cost, and so do not measure what OpenAI actually spends. That caveat is not a footnote. It is the whole interpretive problem with this data.

My position is this. The growth rate is the finding I take from Epoch AI, and the dollar figures are a usage proxy denominated in a price OpenAI charges outsiders. Epoch itself declines to call them expenditure, and its figures are expressed in list-price dollars, not tokens, internal cost or hardware, so nothing in them converts the valuation into a cost or capacity number. So I treat $601 and $7,000 as readings on a usage scale, and I withhold any cost or capacity conclusion until a measurement that can support one appears.

A Yardstick Not a Bill

List price is the unit these figures arrive in, and that is the limit of what I know about the unit. Epoch AI's analysis gives the valuation without a mapping onto tokens or onto compute consumed, and I cannot fill that gap from the source. Epoch AI describes the figures as showing how quickly researchers increased their usage and as illustrating its scale. I take that description at face value and no further. Epoch AI expresses this usage in dollars rather than compute, so I cannot say what hardware, if any, the curve corresponds to, and an operator reading it for capacity would be working from the same gap.

The same discipline applies to the number that will travel furthest. Epoch AI writes that if a 90th-percentile researcher used $7,000 of inference at API prices every working day, the total would come to around $2 million per year, which Epoch calls comparable to or greater than the cost of employing a researcher. That is a conditional illustration built on an every-working-day assumption, and it sets a list-price valuation beside an employment cost Epoch does not itemize. It is not an observed expenditure. I would not carry it into any model of OpenAI's cost base.

What Epoch AI's analysis does not give is any token count, any internal cost per token, or any statement about which party inside OpenAI bears the charge. I am not claiming those numbers do not exist somewhere. I am saying that this analysis does not rest on them, and so neither can mine.

Why the Slope Reads Sticky

Epoch AI's method is straightforward and worth stating. Epoch AI extracted the data embedded in OpenAI's first two charts and fitted trends for the median and 90th-percentile researchers. Its preferred fit is an exponential with a single breakpoint in the slope. For the later months of the window, Epoch AI reports growth of 1.8 times per month for the median and 2.2 times per month for the 90th percentile, which it converts to doubling times of 34 days and 27 days. Both series, Epoch AI says, grew several thousandfold from their first positive January values to mid-August.

Here I lay my own frame over Epoch's numbers, and I want to be explicit that it is mine and not something Epoch measured. I lean toward reading inference consumed inside a research workflow as demand that persists, but the source gives me no premise to rest that on, so I state it as a hypothesis with a test attached. If later data showed the same researchers holding or raising their consumption after the adoption wave passed, the persistence reading would be confirmed. If it showed consumption concentrating in a shrinking group, or falling back toward the January base, it would fail. Epoch's figures also offer no way to check whether the list-price valuation moved in step with the quantity of inference actually consumed, so any capacity reading I take from the slope carries that open question with it.

I should also be honest about what the source does not show. A rising median tells me the middle of the distribution moved. It does not tell me whether the same individuals kept using the tools week after week, or whether a changing population of adopters produced the shape. Stickiness is an inference I am layering on top of a distribution statistic. I think it is the right inference. It is still an inference.

Two Charts One Lab One Breakpoint

The strongest case against my reading starts with the sample. Epoch AI's trends come from two charts in one post by one lab, covering January to mid-August 2026, with the points extracted from those charts by Epoch rather than from a published dataset. The January base was under $1 a day at the median, so a several-thousandfold increase is partly a statement about how small the starting point was. The fit Epoch prefers includes a single breakpoint in the slope, so the later-months rate it reports is one segment of a fitted trend rather than a single rate across the whole window, and nothing Epoch reports says that later slope holds beyond mid-August. Nothing in the data rules out another change. One organization's internal adoption over a window this short is a thin base for any claim about persistence, mine included.

The dollar critique cuts both ways too. Because the series is a list-price valuation rather than a measured expenditure, a flattening in dollars would be as hard to interpret as the rise, and nothing Epoch reports lets me read either direction as a change in compute consumed. Internal usage by OpenAI's own staff also says nothing about paying-customer demand or about margins, because Epoch gives no indication of what, if anything, is charged internally. And if the persistence I am leaning toward turns out to be wrong for this population, the whole capacity-planning implication evaporates with it.

Who Carries the Reading Risk

If this usage persists and is matched by compute actually consumed, the winner is whoever supplies that compute for lab-internal agent workloads. Epoch AI's analysis does not name those suppliers, and I will not guess. The exposed parties are easier to identify. Anyone building an estimate of OpenAI's cost structure or gross margin from a list-price series is exposed, because the series is a valuation, not a cost. Anyone extending the 27-day or 34-day doubling times into a hardware forecast is exposed too, because those are fitted rates over a window Epoch describes with a slope break, and Epoch offers no basis for carrying them forward.

The checkpoint that would falsify me is specific. If a later disclosure of per-researcher usage shows the median and 90th-percentile consumption falling back toward the January base rather than holding near the mid-August levels Epoch reports, then the persistence reading was wrong and I will say so. Slower growth alone would not do it. A high plateau is consistent with my reading. A reversal is not.

Readings That Would Outrank the List Price

Three observables would move my confidence. First, whether OpenAI extends the research-acceleration series beyond mid-August and whether the extension holds near the fitted monthly doubling, slows, or reverses, with a reversal being the result that matters most for persistence. Second, whether OpenAI or Epoch AI publishes any measure of coding-agent use expressed in compute rather than list-price dollars, along with how that measure relates to the dollar series, since Epoch's current analysis offers no such pairing and without one I cannot read the curve in hardware terms. Third, whether OpenAI discloses how much of its inference compute goes to internal coding agents, which, read alongside the first two, would bear on whether this usage is material to its capacity planning.

Until one of those arrives, I have a growth rate I trust and a dollar figure I do not. The curve is the story. The price tag is the unit it happened to be printed in.

Sources

The Current is written by Josh Jiwoon Inn with AI assistance in drafting. It does not claim interviews or first-hand reporting, and every number is attributed to a named source. See the AI-assisted Disclosure and Methodology.