Agents API Public Beta
The engine behind Codex is now open to outside developers.
The autonomous-agent machinery that used to live only inside ChatGPT and Codex is now a public API from OpenAI. Long-running tasks, code execution, and file handling can now be wired directly into any developer's own product.
Codex's insides just
became an API
The core piece — letting an agent reason and act — had never left the building before.
On September 10, 2026, OpenAI launched the Agents API in public beta, exposing the agent execution harness that powers Codex and ChatGPT. Per the official post, Introducing the Agents API, the API is built directly on the open-sourced Codex harness — the reasoning-loop engine itself — so developers can plug in the Model Context Protocol (MCP), their own custom functions, and built-in tools like web search straight into their own agents.
Until now, what OpenAI exposed externally was essentially a single-turn chat API. The part of Codex that lets an agent run autonomously for hours, execute code, and read and write files while completing a task had stayed internal. This release is the first time that line has been crossed.
| Before (Codex-only, private) | Agents API (now) |
|---|---|
| Ran only as internal Codex / ChatGPT machinery | Outside developers can embed it in their own products |
| Basic unit was a single chat turn | Agents run autonomously for hours at a stretch |
| Sandbox environment stayed private | Choose OpenAI-hosted or Cloudflare / Vercel / Oracle |
| Billing model undisclosed (internal use only) | No extra fee — billed purely on token usage |
What early adopters are reporting
Figures OpenAI cited alongside the announcement, also listed in the OpenAI API Changelog.
The Agents API handles automatic context management (the API itself decides what context to keep), runs multiple tool calls in parallel, and can delegate sub-tasks to sub-agents. That's work developers previously had to cover themselves, often by reaching for an external framework like LangGraph or CrewAI — now much of it can be handled by OpenAI's own infrastructure alone.
Billing carries no extra platform fee and is based purely on tokens consumed. Since longer-running agents tend to rack up more tool calls, a token-based model scales with the amount of work done rather than with wall-clock time.
Who this actually changes things for
The biggest shift lands on teams who were already building their own agents.
Engineers
No more building your own sandbox or orchestration layer. MCP support means existing tool integrations carry straight over — at the cost of deeper dependence on OpenAI's own infrastructure.
PMs and the business side
The SafetyKit, Hypha, and Cirridae figures give a concrete starting point for an ROI case. Real running costs at production scale still need your own estimate based on actual usage patterns.
Everyday users
You won't touch this API directly, but more agent-style products built on top of it will start showing up in the tools you already use. The effect reaches you with a delay.
After the chat API, the next thing to go external
was the agent runtime itself.
What to check before you adopt it
This is a beta, so plan for spec changes and breaking updates before betting production workloads on it. Before running long-lived agents in production, three things are worth checking first. First, estimate what token-based billing will actually cost for the task lengths you have in mind. Second, compare OpenAI-hosted sandboxes against the Cloudflare, Vercel, and Oracle partner options against your own data-residency and compliance needs. Third, decide whether this API should replace your existing agent-orchestration tooling outright, or run alongside it for now.
Security is the other thing not to overlook. OpenAI's own Codex agents were tied to a hacking campaign reported in September 2026 that compromised more than ten additional websites. A design where agents act autonomously, for hours, with broad permissions raises productivity — but any gap in sandboxing or permission design becomes an attack surface just as directly. Adopting the Agents API means deciding, at the design stage and with equal weight to the convenience, exactly how much authority an agent should be handed.