Coding agent use at OpenAI is doubling roughly every month

Epoch AI took the charts from OpenAI's September post "Research acceleration: The view inside OpenAI" and fit trend lines to them, and the result is striking. Valued at retail API prices, the median OpenAI researcher's coding agent usage rose from under $1 a day in January 2026 to about $601 a day by mid-August. The 90th-percentile researcher was running past $7,000 a day. In their analysis published on October 5, authors Yafah Edelman and Luke Emberson estimate recent growth at roughly 1.8x per month for the median user, a doubling every 34 days, and faster still at the top of the distribution.

Annualized over 260 working days, that mid-August pace works out to about $160,000 per median researcher and $2 million for the heaviest users. If the recent trend held for another year, the run rate would reach $185 million and $2.3 billion respectively, which the authors flatly call probably unsustainable. The numbers come with caveats: the data is self-reported by OpenAI, API list prices are not what OpenAI pays internally, and the pool of researchers shifts over time, so these are not real expenditures but a measure of how much model work researchers consume.

Even discounted, the shape is clear. The people with the earliest and cheapest access to frontier models are pouring exponentially more compute into their own workflow, and the growth has not visibly leveled off.

Why it matters

If you budget for AI coding tools, this is a preview of where your own bill may head once agentic coding becomes normal: usage priced at API rates can grow faster than any flat seat license assumes. Watch consumption per engineer, not just headcount, and set hard budget caps before the curve finds you.