Epoch AI says the value of coding-agent usage among OpenAI researchers doubled about once a month from January through mid-August 2026. The research group fitted trends to weekly figures that OpenAI had published about its researchers’ agent use.
The result describes the pace of growth during that period. It does not establish that the same doubling rate continued after mid-August.
List-price value is not disclosed spending
Rohan Paul’s summary of the analysis says the median researcher reached $601 per day in coding-agent inference valued at API prices by mid-August, up from under $1 in January.
Those are list-price valuations, not a disclosure of OpenAI’s cash expenses. Paul explicitly noted that the figures do not reveal what the company actually spends. The list-price series therefore cannot establish OpenAI’s internal compute costs, negotiated pricing, or other accounting details.
The doubling rate is not a forecast
Epoch AI researcher Luke Emberson called his own follow-up a rough calculation. Using API prices, he estimated an annualized range of $500 million to $1.5 billion for researchers’ coding-agent usage and said a continued trend would imply tens of billions of dollars per year by the end of 2026.
That extrapolation is conditional, not another measured result. The supported finding is narrower: Epoch AI’s fit shows a steep rise in the API-price value of usage between January and mid-August, while the analysis does not establish that monthly doubling is sustainable.
Comments