OpenAI launches Agents API in public beta

OpenAI launched the Agents API in public beta on 10 September 2026, opening the harness that powers Codex and ChatGPT for Work to all developers through a single API call.

OpenAI said the product lets developers "build and run cloud agents with the Codex harness, fully managed by OpenAI." The launch came the same morning as GPT-Live-1 and ChatGPT for Financial Services, according to MarkTechPost.

The API bundles sessions, orchestration, context compaction, tool search, subagents and recovery into one call that passes the task, model, tools and environment. Context compaction automatically summarises earlier work so agents can keep running across many context windows, OpenAI said. Tool search loads tool definitions only when needed, reducing token use, according to the notes.

An agent built on the API has four parts: the Agent itself, defined by model, instructions, tools and MCP servers; an optional Environment sandbox; a durable Session; and a stream of events. OpenAI said agents can work continuously for hours or days.

Sandboxes can run on OpenAI's own infrastructure, on a customer's cloud, or through partners including Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel. OpenAI runs the harness in every case, regardless of where the sandbox sits.

OpenAI said there are no additional fees for using the Agents API. Customers pay only for model tokens, tools and any OpenAI-hosted container time.

The beta has limits. Data storage is restricted to the US, and Zero Data Retention is not yet supported. Early customers named include Ciridae, Long Lake, WithCoverage, SafetyKit, Dwelly, Hypha, deepsense.ai and Nash.ai.

The launch follows a broader shift in which the harness surrounding a model, not the model alone, increasingly determines what an agent can do. OpenAI's move puts it in direct competition with Anthropic's agent tooling for Claude.