OpenAI’s researchers burned $7,000 a day on AI agents — now it’s opening the floodgates
The New Stack Amanda Caswell ● Covered by 5 sources
OpenAI just opened its Agents API to let developers run AI agents for days. That’s handy, but OpenAI’s own researchers already showed how fast agent compute can get expensive.
Based on reporting by The New Stack, Amanda Caswell — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI has put its Agents API into public beta, exposing the backend behind Codex to developers who want to run agents unattended for long stretches. The pitch is simple: developers no longer have to build their own orchestration stack just to keep a task alive past a single context window. The API keeps track of where the job is, gives it somewhere to run, and lets it keep moving.
That convenience comes with a bill. As tasks get longer, the API can compress earlier context instead of stopping at the model limit. It can pull in tools only when they’re needed, and it can split bigger jobs across subagents working in parallel. The work can happen in OpenAI’s sandbox or on infrastructure the developer controls. Either way, each step sends the agent back to the model, which means hours of work can rack up a lot more inference than a normal API call.
OpenAI has already seen the shape of that problem inside its own research team. In a report published September 6, the company said its research organization was logging 3.1 agent-workdays for every human workday by mid-August, using standard eight-hour equivalents. The median researcher, ranked by agent usage, was spending more than $600 a day on inference at API prices. The 90th percentile was over $7,000. Before June, people were still doing more work than their agents. By mid-August, the agents were doing three times as much.
The timing makes the rollout feel even more pointed. On the same day the Agents API launched, OpenAI paused new sign-ups for its $200-a-month Pro plan after demand for GPT-6 Astra strained capacity. Thibault Sottiaux, who leads engineering for Codex, said on X that Pro subscriptions were putting the most strain on OpenAI’s systems and that the company was adding capacity as fast as it could. The Agents API and ChatGPT Pro are separate products, so one isn’t directly stealing resources from the other. Still, it’s a neat snapshot of where the pain is: OpenAI wants developers to run agents for hours or days, while also having to slow access to its heaviest-use consumer plan.
The bigger lesson is that agent usage scales in ugly, expensive ways. One person can light up far more compute than headcount suggests, especially when several agents are running in parallel. OpenAI’s own numbers make that painfully clear.
My take — AI-written commentary, not fact-checked reporting
This is the AI industry’s favorite trick: make the thing easier to use, then act surprised when the bill grows legs. OpenAI isn’t just selling agents here; it’s normalizing a world where compute spend, not headcount, becomes the real measure of activity. Very efficient, if the goal is to turn every workflow into a small cloud invoice.
Read more about this at: The New Stack