OpenAI researchers now spend $600 a day on AI tools
A post from inside the lab shows how far OpenAI has handed its own research work to coding agents, and what that daily habit now costs.
Symbolic image: a workstation during a nighttime agent run, status lights blinking on the machine while an unlabeled curve climbs across the screens.
OpenAI's own account of its research acceleration shows daily AI-tool spend per researcher rising from near zero in February 2026 to about $600 by late August 2026, as staff moved their work onto coding agents.
At a glance
- Daily AI-tool spend per researcher: near zero until February 2026, about $150 in June, about $600 by late August 2026.
- The figures come from a chart in OpenAI's post, as read and described by Simon Willison on September 6, 2026.
- OpenAI's own framing is that 2026 is the year agentic engineering really took off inside the company.
- Willison guesses the late-July jump lines up with internal access to the model later shipped as GPT-6 Astra.
- OpenAI's page returned HTTP 403 during writing, so the details here rest on Willison's summary of it.
The most concrete thing in OpenAI's look at its own research work is a spending line. Daily AI-tool spend per researcher sat near zero until February 2026, reached roughly $150 in June, and climbed to about $600 by late August 2026.
The chart is the argument
Those numbers are plotted, not tabulated. Simon Willison read the chart on September 6, 2026 and traced its shape: flat through February, a gradual rise into early summer, then a steep climb from late July onward. The interesting part is not the money but what the money buys — a change in how the work gets done.
Agents as the default way to work
OpenAI describes 2026 as the year “agentic engineering really took off” in the company, and publishes the piece alongside an essay from chief scientist Jakub Pachocki. Willison reads both as being about recursive self-improvement, which he calls the lab's “new AGI” — his phrase, not OpenAI's.
What happened in late July
Neither text, on Willison's reading, explains the inflection. His own guess is that staff got internal access around then to the model that later shipped as GPT-6 Astra. He flags it as a guess, and OpenAI has not confirmed it.
What we could not verify
The OpenAI page itself returned HTTP 403 while this article was being written, so it could not be read directly. Every figure above comes from Willison's description of the chart. Two things stay open: what OpenAI counts as tool spend — inference, licenses, or both — and whether the line covers all research staff or one slice of them.
Why an outside reader should care
For a lab this size, $600 per person per day is a choice rather than a strain. It says agent runs are now budgeted the way compute has always been budgeted. Teams elsewhere will not match that number, but it sets a reference point: this is how much tooling the company building the models considers worth paying for.
FAQ
What is in OpenAI's research acceleration post?
It describes how coding agents became the normal way OpenAI's research staff work during 2026, illustrated with a chart of daily AI-tool spend per researcher.
How much do OpenAI researchers spend per day on AI tools?
Near zero until February 2026, about $150 in June, and about $600 by late August 2026, according to the chart as described by Simon Willison on September 6, 2026.
Did GPT-6 Astra cause the late-July spending jump?
That is Willison's speculation, not a confirmed link. In his account of the post, OpenAI gives no cause for the steep rise.
