Rapport de veille

OpenAI says its researchers' theoretical agent token costs now exceed their base pay

A second internal OpenAI report tracks rising use of coding agents in model development, while staff interviews point to weak results on complex engineering work.

INSIGHTEX Editorial Desk

2 min de lecture

OpenAI says its researchers' theoretical agent token costs now exceed their base payRapport de veille

What happened

OpenAI has released a second report measuring how AI agents speed up its own research, The Decoder reported. The company says the goal is to document signs of recursive self-improvement, where AI systems accelerate work on their successors. Priced at API list rates, the median researcher used tokens worth $2,400 a day in October, a 62 percent rise from the prior month. For the heaviest 10 percent of users, the figure doubled to $23,000 daily.

Between March and October, OpenAI estimates that the success rate of agents working unattended on two- to four-hour tasks climbed from 23 to 46 percent, and from 13 to 35 percent on four- to eight-hour tasks. In 79 percent of cases, agents successfully kept watch over training runs. They succeeded only 29 percent of the time when building evaluation and reinforcement-learning environments, and 25 percent of the time when writing inference and serving code.

In a non-representative internal survey of 101 employees, the median researcher said they would sooner give up 40 percent of team headcount or 30 percent of compute than work three months without current coding agents. Interviews with ten researchers described the downsides: the models show poor judgment about research ideas and code quality, take needless steps, and submit low-grade code changes. The researchers said close human supervision is still needed.

Context

OpenAI's earlier report on research acceleration, released in early September, stated the company had reached a self-set target of an automated AI research intern. In an accompanying post, chief scientist Jakub Pachocki cautioned about losing control after an audit found gaps in oversight. Separately, METR, an evaluation group, estimated after eleven weeks inside Anthropic that AI had almost doubled its development speed.

SourceThe Decoder

Le briefing

La semaine tech, en un seul e-mail.

Les dépêches qui méritent votre temps, chaque semaine. Désinscription à tout moment.