OpenAI: Over $600 a Day in Tokens for Coding

OpenAI reports that by mid-August, the median researcher on its research team was spending over $600 per day on tokens for coding agents, compared to $162 in July. Users in the 90th percentile exceeded $7,000 daily, based on API pricing estimates. Meanwhile, the company states it has reduced internal technical support requests by more than half since January and aims to have a fully automated AI researcher by March 2028. In the ecosystem, players like Meta and Amazon have shut down internal rankings following instances of "tokenmaxxing."

Meta and Amazon Cut Rankings, OpenAI Claims Extreme Usage

Several companies initially encouraged intensive experimentation with AI by establishing internal user rankings. This incentive fostered "tokenmaxxing," which involves pushing the AI to its limits without considering the results. By summer, the high costs of leading models raised doubts about the productivity achieved relative to expenses, leading both Meta and Amazon to close these rankings. At Amazon, Senior Vice President Dave Treadwell urged against "using AI just for the sake of using AI."
OpenAI, for its part, has already highlighted very high consumption levels. In June, Sam Altman indicated that an internal user was reaching 100 billion tokens per month. Peter Steinberger, who was hired in February, shared a screenshot in May of a monthly bill totaling $1.3 million, clarifying that he was not responsible for it, like other employees.

At OpenAI, the Median Exceeds $600 Per Day, the 90th Percentile Over $7,000

OpenAI indicates that its researchers are now using AI coding agents at a costly rate. By mid-August, the median researcher in the research organization was consuming over $600 per day in tokens, up from $162 in July, and users in the 90th percentile were exceeding $7,000 daily. The company specifies that these amounts are based on estimates from API rates and do not reflect actual payments. The most active profiles surpass the $7,000 per day threshold, according to a post published on Sunday. This intensification over the past three months is occurring at a faster pace than in other teams, amid industry debate surrounding "tokenmaxxing."

Internal Support Down and Target Set for an AI Researcher by March 2028

OpenAI observes that the use of agents to solve technical problems is accompanied by a drop of more than half, since January, in messages in a human-managed internal support channel. The company also describes a year in which its researchers are delivering code faster, multiplying experiments, and delegating more complex tasks to agents. It announces that it has achieved the goal set last year by Sam Altman of an "automated research intern" completed under human supervision, and claims to be on track for a fully automated AI researcher by March 2028.
In software engineering broadly, many developers are already delegating coding tasks to agents and verifying their outputs. Some report increased productivity, while others do not, and the practice has proven costly for certain organizations.