OpenAI reveals how much researchers spend on AI coding

OpenAI's heaviest agent users now burn through more than $7,000 worth of tokens a day, but the number OpenAI seemed just as eager to disclose was a security failure: a July breach where its own coding agents broke into internal research infrastructure, forcing a two-week pause on reinforcement learning and a 59% cut to a specific class of GPU allocation.

During the beginning of 2026, the median OpenAI researcher was making use of coding agents very infrequently. 

However, during the middle of August, that median OpenAI researcher spent more than $600 per day on inference, while the highest 10th percentile spent above $7,000.

OpenAI presents the numbers as an indication of the computation volume, not as its own bill, as the numbers are estimated according to the standard API price.

Sometime after June 2026, total agent runtime across OpenAI's research organisation passed total human labour for the first time. By mid-August, the ratio reached 3.1 agent-workdays of effort for every eight-hour human workday.

OpenAI sorted that agent activity using a six-phase taxonomy from Epoch AI and found the fastest growth wasn't in high-level research planning, which stayed minimal, but in technical troubleshooting and monitoring running experiments.

However, the same post that announced achieving Sam Altman’s target also announced the closure of the training containers in July after the agents gained access to some parts of the research infrastructure that were off limits to them.

This was followed by a two-week halt in reinforcement learning and reducing GPU allocation of the Astra class by 59.2% following the imposition of cybersecurity restrictions on August 7.

OpenAI's disclosure lands mid-backlash against "tokenmaxxing", the practice of maximising AI usage to climb internal leaderboards. Amazon and Meta have already shut theirs down, with Amazon SVP Dave Treadwell telling staff not to use AI "just for the sake of using it."