OpenAI has pulled back the curtain on how much its researchers spend on AI coding agents.
In a blog post on Sunday, the startup said the most prolific users of AI coding agents among its researchers now consume more than $7,000 worth of tokens — the small chunks of text an AI model reads and produces — a day.
The technology has substantially changed the way OpenAI’s researchers work over the past year, the company said, with staff shipping code faster, performing more experiments, and delegating increasingly high-level tasks to coding agents.
OpenAI said that AI coding agent usage among its researchers has skyrocketed over the past three months and is outpacing growth among other OpenAI teams.
As of mid-August, the median researcher at the company’s research organization was using more than $600 in tokens per day, up from $162 in July, per OpenAI data, with 90th-percentile users burning through over $7,000 per day. The figures are based on inference at API prices, meaning they’re estimates of what they would cost rather than the dollar amount OpenAI actually paid.
OpenAI also said that its agents were increasingly being used to troubleshoot technical issues, so much so that the number of daily posts in a human-staffed internal technical support channel dropped by more than half since January.
The company said it had reached the goal — set by CEO Sam Altman last year — of building an “automated research intern” that can carry out tasks with human direction, and was on track to create a fully automated AI researcher by March 2028.
Coding’s great reset
The growing popularity and proficiency of coding agents have radically transformed software engineering over the past year.
Many developers now delegate their coding tasks to AI agents and review their outputs. Some say this makes them more productive, but others are less on board.
It has also proved expensive for some companies.
Earlier this year, companies encouraged employees to experiment with AI, including with leaderboards to track the power users. That led some workers to engage in “tokenmaxxing,” the idea of using AI as much as possible, regardless of output.
By the summer, however, the soaring cost of top AI models began to fuel concerns that tokenmaxxing was not increasing productivity enough to justify the spending.
Companies including Meta and Amazon shut down internal token leaderboards, with Amazon senior vice president Dave Treadwell warning staff not to “use AI just for the sake of using AI.”
OpenAI has previously embraced high token usage. Altman said in June that the company’s highest token user burns through 100 billion tokens a month.
OpenClaw creator Peter Steinberger, who was hired by OpenAI in February, posted a screenshot of a $1.3 million monthly token bill in May — although like other employees, he doesn’t have to foot the bill.
