OpenAI Agents API Public Beta: Managed Runtime, Token-Only Billing

The OpenAI Agents API entered public beta on September 10, 2026, per Tech Insider's tutorial.

Quick answer for the scanners: no platform fee, beta live since September 10, 2026. And the only cost is the tokens plus tools your agents use.

What shipped under that date matters more than the date itself.

The Codex runtime turned into a service you call instead of a loop you build and babysit yourself.

The API changelog describes it as building agents with a managed Codex runtime "while OpenAI handles session orchestration, context compaction, and recovery."

Read that quoted list twice.

Orchestration, compaction, recovery.

The three failure modes that page you at 2am. Now sit in somebody else's on-call rotation.

That transfer, not the word "agents," is the actual news.

What the OpenAI Agents API Actually Is

OpenAI's agents guide frames the product as building and running cloud agents on the Codex runtime, fully managed by OpenAI. Same runtime, same infrastructure that powers Codex, handed over through a single API.

OpenAI also publishes the runtime as open source, which leaves the core logic coordinating model calls, tools. And context readable rather than sealed in a box.

The open-source detail earns more weight than one sentence suggests. An agent starts burning tokens on something stupid, and a readable runtime means the answer is a search away. Opaque runtime means a support ticket and a wait.

Surface area stays deliberately small. AIBase reports one core endpoint: `POST https://api.openai.com/v1/agents/sessions`.

Every request carries the `OpenAI-Beta: agents=v1` header. The quickstart says the OpenAI SDKs attach it for you, while raw cURL means typing it yourself. A typical request creates a session, submits a task. And streams progress back, with SDK examples living in the `beta.agents` namespace. Session in, task in, events out. That is the whole shape of the thing.

How the OpenAI Agents API Differs from the Agents SDK

Three entry points now sit in OpenAI's guide: the Responses API, the Agents SDK, and the Agents API.

The guide defines agents as things that "plan and complete tasks using tools, work with other agents. And maintain context across steps," and each of the three products takes a different slice of that sentence.

SDK route means self-service.

The SDK quickstart is `npm install @openai/agents zod` or `pip install openai-agents`, set `OPENAI_API_KEY`, go.

The SDK handles the model call and hands back a result object holding the final output plus the run history.

Process, retries, compaction — those stay on your side of the fence. OpenAI's developer blog positions the updated SDK for agents that "inspect files, run commands, edit code. And work on long-horizon tasks within controlled sandbox environments." Real capability. Your hosting bill.

The Agents API inverts the ownership.

Same stack, OpenAI's cloud runs the runtime, tasks arrive over REST.

Who owns the run that dies at 2am?

OpenAI does, now. For a small team the choice has almost nothing to do with feature lists. It is a decision about whose pager goes off.

The managed option quietly took the three most annoying failure modes, session state drifting, context overflowing, runs dying mid-task, onto OpenAI's side of the ledger.

OpenAI Agents API Pricing: One Line, Then the Trap

The announcement kills the money question in a single line: "There are no additional fees for using the Agents API – you simply pay for the tokens and tools your agents use, as outlined on our pricing page."

No surcharge stacked on top of usage.

For a lean shop, that structure is correct.

A workflow that stops earning can be switched off, no contract looming over the exit. Price of the experiment equals price of the tokens. Freight runs on the same logic — empty truck, small bill — except here there is no truck payment at all, just the meter.

Now the catch, the one that catches every operator eventually. A managed loop retries and compacts on your behalf. Every one of those actions runs on tokens with your name on the invoice, so convenience multiplies spend. The `agents=v1` header is as well a reminder that this is v1 of a public beta. Budget for the interface shifting underneath you before anything gets welded into a client deliverable.

Standing rule for any new API: small test project, watched spend, nothing production-critical until it survives a real workload.

OpenAI Agents API FAQ

Is the OpenAI Agents API free?

Free to call, yes. Free to run, no. The exact wording from the announcement: "There are no additional fees for using the Agents API – you simply pay for the tokens and tools your agents use, as outlined on our pricing page."

What is the Agents API endpoint?

`POST https://api.openai.com/v1/agents/sessions`, per AIBase. Every request needs the `OpenAI-Beta: agents=v1` header.

When did the public beta launch?

September 10, 2026.

One Evening, Three Decisions

Start with the quickstart, since OpenAI scoped it well for evaluation. It builds a coding assistant that writes `tree.py`, runs it, and prints a directory tree. Write, execute, report. A complete loop in an afternoon, roughly one evening and a rounding error in tokens.

It too answers whether the streaming model fits how you like to work.

Next, leave working Agents SDK code alone on day one.

A stable loop under your own hosting is still a legitimate path. And OpenAI's own docs present it that way rather than as a dead end. The Agents API earns its slot when the runtime is the part you want off your plate. Not because it is the newest name in the docs.

Last decision belongs on paper before anything ships: the handover list.

Session orchestration, context compaction, and recovery are OpenAI's problem now, and that is genuinely valuable. Permissions, testing, and what an agent may touch inside a client's systems remain yours. The day something breaks is the wrong day to learn you assumed OpenAI was checking something it never promised to check.

The public beta is OpenAI selling the plumbing instead of the promise.

Small operators should spend one evening on it and probably adopt it for the right workload.

A second pair of eyes on whether your automation belongs on a managed runtime or stays in your own stack. That is precisely the call Mediascout helps builders make.

Sources

- Tech Insider tutorial - OpenAI API changelog - OpenAI agents guide - AIBase report - Agents API quickstart - Agents SDK quickstart