OpenAI reveals how much its researchers are spending on AI coding

OpenAI has pulled backmost the curtain connected really overmuch its researchers walk connected AI coding agents.

In a blog station connected Sunday, the startup said the astir prolific users of AI coding agents among its researchers now devour much than $7,000 worthy of tokens — the mini chunks of matter an AI exemplary sounds and produces — a day.

The exertion has substantially changed the measurement OpenAI's researchers activity complete the past year, the institution said, pinch unit shipping codification faster, performing much experiments, and delegating progressively high-level tasks to coding agents.

OpenAI said that AI coding supplier usage among its researchers has skyrocketed complete the past 3 months and is outpacing maturation among different OpenAI teams.

As of mid-August, the median interrogator astatine the company's investigation statement was utilizing much than $600 successful tokens per day, up from $162 successful July, per OpenAI data, pinch 90th-percentile users burning done complete $7,000 per day. The figures are based connected conclusion astatine API prices, meaning they're estimates of what they would costs alternatively than the dollar magnitude OpenAI really paid.

OpenAI besides said that its agents were progressively being utilized to troubleshoot method issues, truthful overmuch truthful that the number of regular posts successful a human-staffed soul method support transmission dropped by much than half since January.

The institution said it had reached the extremity — group by CEO Sam Altman past twelvemonth — of building an "automated investigation intern" that tin transportation retired tasks pinch quality direction, and was connected way to create a afloat automated AI interrogator by March 2028.

Coding's awesome reset

The increasing fame and proficiency of coding agents person radically transformed package engineering complete the past year.

Many developers now delegate their coding tasks to AI agents and reappraisal their outputs. Some opportunity this makes them much productive, but others are little connected board.

It has besides proved costly for immoderate companies.

Earlier this year, companies encouraged labor to research pinch AI, including pinch leaderboards to way the powerfulness users. That led immoderate workers to prosecute successful "tokenmaxxing," the thought of utilizing AI arsenic overmuch arsenic possible, sloppy of output.

By the summer, however, the soaring costs of apical AI models began to substance concerns that tokenmaxxing was not expanding productivity capable to warrant the spending.

Companies including Meta and Amazon unopen down internal token leaderboards, pinch Amazon elder vice president Dave Treadwell informing unit not to "use AI conscionable for the liking of utilizing AI."

OpenAI has antecedently embraced precocious token usage. Altman said successful June that the company's highest token personification burns done 100 cardinal tokens a month.

OpenClaw creator Peter Steinberger, who was hired by OpenAI successful February, posted a screenshot of a $1.3 cardinal monthly token measure successful May — though for illustration different employees, he doesn't person to ft the bill.