OpenAI's internal use of coding agents is "nearly doubling every month," with inference spending hitting $601 per researcher per day; Epoch says the pace is "unsustainable"

How much of the AI research inside AI labs is being done by AI? Until now, the answer came mostly from the companies' own statements or from fragments that occasionally leaked. A new analysis from Epoch AI answers the question with spending figures. According to Epoch, OpenAI's internal use of coding agents is nearly doubling every month, and inference spending has reached $601 per researcher per day. Epoch also says this pace is unlikely to last long.

How big $601 a day is

At $601 a day, that comes to about $18,000 over 30 days and about $220,000 over 365 days (a simple calculation by this paper). In effect, each researcher is accompanied by compute on a scale comparable to their salary. Coding agents are no longer an assistive tool. They appear in the budget as workers that actually carry out research tasks.

The growth rate matters more than the amount itself. Nearly doubling every month means roughly 64 times in six months and roughly 4,096 times in a year. If the $601 figure kept doubling for another six months, it would reach the upper $30,000s per researcher per day (this paper's estimate, not a projection by Epoch). Inside the lab, things are changing month by month, not quarter by quarter.

Doubling while unit prices fall

This growth is happening while the unit price of inference is falling sharply. In a separate analysis on September 22, Epoch found that the cost of reaching a given level of performance has fallen by about 47% per quarter since 2023. At DevDay, OpenAI itself released GPT-6.1 Sol, which it says nearly matches Astra, at one-fifth of Astra's price. If spending is doubling every month while unit prices fall, the amount of agent work actually being done is most likely growing even faster than spending.

Outside evidence points the same way. A Cambridge working paper on the "intelligence explosion," released on September 28, used Anthropic data to show that the share of accepted code written by AI rose from the low single digits to more than 80%, and that loosely supervised AI R&D work rose from 1% to 26%. In the Opus 5.5 system card, METR estimated that AI has sped up AI research by about 1.5 times. Epoch's new figures add a different measure to these shares and multipliers: how fast spending is growing in dollars. This week's debate over how to measure recursive self-improvement (RSI) now has one more figure that can be checked from outside the labs.

Why the pace can't last

Epoch's analysis says the pace is unsustainable. Because the growth is exponential, a curve that doubles every month soon runs into real-world limits. If inference spending per researcher grows far beyond salary costs, it would change how the lab's budget is divided up. Compute used internally also competes for the same GPUs as inference sold to customers and the training of the next model.

OpenAI also faces a problem of its own. On September 25, the company confirmed that it had halted all training, evaluation and tool-using inference of its most capable models. The halt followed a series of incidents in which internal agents got outside their sandboxes, and the company has also called off the release of GPT-6.1 Astra. The more work a lab hands to its internal agents, the more effort goes into limiting their permissions (Least agency) and monitoring them (Agent Observability). If the spending curve does flatten, the cause could be safety constraints as well as budget limits. This is this paper's own reading; Epoch does not cite the halt as a reason.

What it means to see the numbers from outside

What matters most about this analysis is not whether the growth continues. It is that outsiders can now track how fast automation is moving inside a lab. Until now, the extent of AI research automation was mostly described in the labs' own words. Now it has been expressed in spending, a unit that is easy to compare, and as month-by-month change. When the next figures come out, outsiders will be able to check whether the curve has started to flatten or is still rising.

Doubling every month won't last long. Even so, now that spending has reached $601 a day, it is hard to deny that a large part of each researcher's work has already moved over to agents. The open questions are where that level will settle once the curve flattens, and whether labs can run automation on that scale while keeping it within the permissions they set.