OpenAI turns its internal Codex agent harness into a billable API primitive
The new Agents API exposes hosted orchestration, long-running sessions, sandboxed code execution, and multi-agent delegation directly, so cost now scales with sandbox and container usage as well as tokens.
OpenAI opened public beta of its Agents API on September 10, 2026, exposing the managed harness that powers Codex as a direct API call: hosted orchestration, long-running sessions, context compaction, sandboxed code execution, MCP connections, and multi-agent or subagent delegation. It's priced at standard token and tool costs with no added markup.
Infrastructure becoming a product
Per OpenAI's developer documentation, the Agents API "gives your application access to the Codex harness through an OpenAI-managed API." Internal infrastructure that supported one product is now a general-purpose building block any developer can call.
A new line on the spend forecast
If you forecast agent-tooling spend, cost now scales with sandbox and container usage in addition to tokens. OpenAI's developer community flagged this directly, with one member publicly cautioning others to model the costs before building on it. Finance teams that track token spend as the main AI-cost line need a second line for sandboxed compute time.