OpenAI's Agents API Enters Public Beta: What It Actually Does

OpenAI opened its Agents API to public beta on September 10, 2026, handing every developer the same orchestration engine that runs Codex — session management, context handling, and tool coordination — behind a single API call instead of custom-built infrastructure. That's the actual news: not a new model, but OpenAI packaging the hardest part of building an AI agent (the plumbing, not the prompting) and shipping it as a product.

If you've ever built an agent that has to remember what it did ten steps ago, call tools in the right order, and not fall over when a session runs long, you know that plumbing is most of the work. OpenAI is now selling you its own, already-battle-tested version of it.

Quick Take

• What it is: A public beta API that exposes the same harness powering OpenAI's Codex coding agent — session handling, context management, and tool orchestration — for developers to build any kind of agent, not just coding ones.
• Pricing: No separate platform fee. You pay for the tokens, tools, and compute your agent actually uses, the same as any other OpenAI API usage.
• Where it runs: In an OpenAI-managed sandbox, on your own infrastructure, or through a first-party integration with partners including Cloudflare, DigitalOcean, Oracle, and Vercel.
• The catch: It's a beta, not a finished product — OpenAI says it's still iterating toward general availability, which means interfaces and defaults can still shift under you.

What's Actually New Here

Every team building an AI agent runs into the same unglamorous problems: how do you keep a session's context from blowing past the model's window, how do you decide which of your fifty tool definitions to actually show the model on a given turn, and how do you let several subagents work in parallel without stepping on each other. None of that is prompt engineering — it's infrastructure, and most teams have been writing their own version of it by hand.

The Agents API is OpenAI's answer: it hands you the same harness that already runs Codex, its coding agent, and lets you point it at your own tools and use cases instead. According to reporting on OpenAI's announcement, the harness handles automatic context compaction as a session nears its limit, "tool search" that loads tool definitions only when the model actually needs them instead of stuffing all of them into every prompt, programmatic tool calling, and native support for running subagents in parallel. Your application still controls the interface, the actual tools, and the approval flow — OpenAI's harness runs the agent loop underneath it.

That harness is also open source, which is the part worth sitting with. OpenAI isn't just renting you a black box; it's shipping the same code Codex runs on and letting you inspect or extend it. For teams who got burned trusting closed agent frameworks before, that's a meaningfully different pitch than "trust our API."

Pricing and Where Agents Actually Run

There's no new subscription tier or per-agent fee bolted on top. You pay for tokens, tool calls, and whatever compute your agent's sandbox consumes — the same metered pricing OpenAI already applies to the rest of its API, layered onto whichever model you point the agent at.

The more interesting decision is where the agent's sandbox actually lives. OpenAI gives you three options: run it inside an OpenAI-managed sandbox with nothing to provision, run it on your own infrastructure, or use one of its launch sandbox partners — Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel all have first-class integrations at launch. That's a notably broad list for a day-one beta, and it signals OpenAI wants the Agents API embedded in infrastructure developers already use rather than pulling everything into its own cloud.

Why This Is the Sequel to the Assistants API Shutdown

If this sounds familiar, it should. OpenAI just shut down its Assistants API on August 26, 2026, after a year-long deprecation notice, pushing everyone still building stateful agents onto the Responses API. The Agents API is the next layer up that migration: where Responses gave developers a lower-level, stateless building block, the Agents API is the opinionated orchestration layer built on top of it, wrapping the same kind of session and context problems the Assistants API tried (and, by OpenAI's own admission, struggled) to solve.

It's also a direct answer to a gap that's been growing between OpenAI and its rivals. Anthropic has spent 2026 building out its own Claude Agent SDK — the renamed successor to the Claude Code SDK — and has pushed Model Context Protocol, the connector standard behind its agents, through a major stateless spec rewrite this summer specifically to make agent infrastructure easier to run at scale. Google has been doing the same with its open-source Agent Development Kit. All three companies have landed on the same conclusion at roughly the same time: the model isn't the bottleneck anymore, the orchestration layer around it is, and whoever owns that layer keeps developers on their platform.

How It Fits Next to Claude Code and Cursor

For teams already choosing between agentic coding tools like Claude Code, Cursor, and GitHub Copilot, the Agents API isn't really a competitor to any of them — it's a layer underneath, meant for developers building their own agents rather than buying a finished one. If you use Codex directly today, this doesn't replace it; it's the same engine Codex runs on, now available for anything else you'd want an agent to do, from an internal ops dashboard to a customer-facing support bot.

The practical tradeoff is the same one every "we handle the hard infrastructure part" pitch carries: you get faster time-to-build and less orchestration code to maintain, in exchange for depending on OpenAI's opinions about how context should be compacted and when tools should be surfaced. For a straightforward internal agent, that trade is probably worth it. For something with unusual session or compliance requirements, the self-hosted-infrastructure option is there specifically so you're not locked into OpenAI's managed sandbox.

Frequently Asked Questions

Is the OpenAI Agents API free to use? There's no separate subscription for the API itself. You pay standard OpenAI usage rates for the tokens, tool calls, and compute your agents consume, the same as any other API workload.

Do I need to use OpenAI's own sandbox to run agents? No. You can run an agent's compute in an OpenAI-managed sandbox, on your own infrastructure, or through a launch partner such as Cloudflare, DigitalOcean, Modal, or Vercel.

How is this different from the Assistants API OpenAI just shut down? The Assistants API managed threads and tool calls directly and was retired in favor of the lower-level Responses API. The Agents API sits a layer above Responses, adding the orchestration — context compaction, tool search, parallel subagents — that the Assistants API tried to provide but that OpenAI has now rebuilt on newer foundations.

Is the underlying harness open source? Yes. The Agents API is built on OpenAI's open-source Codex harness, so developers can inspect the orchestration logic rather than relying on a fully closed system.

Is this ready for production today? It's a public beta, not a general-availability release. OpenAI says it's actively iterating based on developer feedback, so treat defaults and interfaces as subject to change until GA lands.

Bottom Line

The Agents API isn't a flashier model or a new consumer feature — it's OpenAI finally packaging the orchestration work every serious agent builder was already doing by hand, and open-sourcing the harness that runs it. If you're building an agent today, it's worth a serious look before you write your own context-management layer from scratch; just remember it's beta software from a company that only just finished forcing a hard migration off its last agent API.