OpenAI Launches Agents API in Public Beta, Exposing Codex Infrastructure to Developers

OpenAI released the Agents API in public beta on September 11, 2026, giving developers direct access to the cloud infrastructure that already runs Codex and ChatGPT's agentic features.

The API enables developers to build cloud-based agents capable of running for hours at a stretch, executing code, and processing files—capabilities previously locked inside OpenAI's own products.

What the API Does

According to OpenAI, the Agents API includes three core capabilities:

• Automatic context management — the system handles context window limits without requiring manual intervention from developers
• Parallel tool use — agents can call multiple tools simultaneously rather than sequentially
• Sub-agent delegation — a primary agent can spin off tasks to secondary agents to handle in parallel

Developers get a choice of execution environments: OpenAI-hosted sandboxes, or third-party infrastructure from Cloudflare, Vercel, and Oracle. This multi-cloud approach lets teams run agent workloads on infrastructure they already use or trust for compliance reasons.

Pricing and Technical Foundation

OpenAI says there are no additional fees for using the Agents API beyond standard token-based billing—meaning the extended runtime, sandboxing, and orchestration features come at no separate cost. OpenAI has not disclosed specific per-token pricing for this API in the announcement.

The API is built on top of the open-source Codex harness, the same underlying framework used in OpenAI's coding agent product. It supports the Model Context Protocol (MCP), custom function calling, and built-in tools including web search.

What This Means

This is not a new model release—it's OpenAI productizing internal infrastructure. By exposing the Codex/ChatGPT agent runtime as a standalone API, OpenAI is competing directly with a growing field of agent orchestration frameworks and cloud providers offering similar capabilities, including LangChain, CrewAI, and cloud-native agent tooling from AWS and Google.

The zero-extra-fee pricing structure is a notable competitive move: rather than charging separately for compute-heavy features like hours-long execution or sandboxing, OpenAI is betting that token volume alone will capture the economics. This could pressure competitors who monetize agent infrastructure through separate compute or orchestration fees.

The multi-cloud sandbox options (Cloudflare, Vercel, Oracle) also signal OpenAI's strategy of meeting enterprise developers where they already deploy infrastructure, rather than requiring a full migration to OpenAI's own hosting. Given that this API builds on the same system already powering Codex, expect close feature parity between what ships in OpenAI's own agent products and what third-party developers can now build themselves.