OpenAI’s researchers burned $7,000 a day on AI agents — now it’s opening the floodgates

OpenAI launched its Agents API in public beta on Thursday, enabling developers to run agents unattended for extended periods without building their own infrastructure. This rollout coincided with OpenAI pausing new sign-ups for its $200-a-month Pro plan due to demand for GPT-6 Astra pushing capacity limits. While the two products are separate, the timing is notable as OpenAI is simplifying agents' long-term operation while simultaneously restricting access to its most resource-intensive consumer plan.

Agent inference can become costly over time, with powered-by-the-API tasks potentially racking up significantly more usage than a regular API call, especially when agents operate in parallel. OpenAI's own research team logged 3.1 agent-workdays for every human workday by mid-August, with the 90th percentile of researchers spending over $7,000 daily on inference at API prices.

This demonstrates how agents can generate far more inference than a single human, raising concerns about compute demand. OpenAI's approach to agent context compaction reduces friction for developers by automatically handling earlier context compression, but also means work can continue even when the context window fills up.