OpenAI’s Agents API, the technology behind its Codex agent, is now available in public beta, allowing developers to build and run autonomous agents without managing their own orchestration layer. The API handles context compaction, letting tasks stretch beyond a single context window by compressing earlier work, and can invoke tools on demand or dispatch subagents to process parts of a job in parallel. Developers can execute agents inside OpenAI’s sandbox or on their own infrastructure.
While the move lowers the barrier for creating long-running agents, it also lowers friction for consuming inference. OpenAI’s own research organization provides a vivid example of how fast usage can scale. In a report published September 6, the company said its researchers logged the equivalent of 3.1 agent-workdays for every human workday by mid-August. The median researcher spent more than $600 per day on inference at API prices, and the 90th percentile exceeded $7,000 per day. Before June, researchers were still logging more hours than their agents; by mid-August, agents were doing three times the work of their human counterparts.
The timing of the launch coincides with a decision to halt new Pro subscriptions. Thibault Sottiaux, engineering lead for Codex, said on X that Pro plans “put the most strain on our systems” and that OpenAI was adding capacity as quickly as possible. The two products are separate, but both draw from the same underlying inference infrastructure.