OpenAI's Agents API Is Free. That's the Strategy.

On Wednesday OpenAI put the machinery that runs Codex behind a public API and declined to charge for it. The Agents API entered public beta with a pricing line you rarely see from this company: no additional fees. You pay for tokens, tools, and container time. The orchestration around them comes free.

I build with agents every day, and I know exactly which parts of my stack this replaces, because I wrote them myself. Session recovery after a crash. Context that outgrows the window halfway through a task. Subagent fan-out when a job splits cleanly. That code is boring, fragile, and never finished. OpenAI now offers to run it for me at no charge, and the interesting question is what a company with OpenAI's margins expects to collect later.

What actually shipped

The announcement describes a managed orchestration layer, the same one behind Codex and ChatGPT for Work, exposed over an API. Sessions persist and can be resumed. You can steer an agent mid-turn. Work can be split across subagents with a configurable concurrency cap. For the environment, you get three options: an OpenAI-hosted sandbox billed at standard container rates, a self-hosted executor that runs inside your own infrastructure, or an integration with one of nine launch partners, including Cloudflare, Modal, E2B, Oracle, and Vercel.

The post leans on customer numbers. One team reports an evaluation score going from 0.71 to 0.85 with a 4x latency reduction from subagent flows. Another reports 60% lower cost per case after migrating a review workflow. These are OpenAI-selected quotes, not independent benchmarks, so I read them as directional. Even discounted, they point at real orchestration pain that teams are currently solving by hand.

The parts you don't pay for are the parts you stop owning

The self-hosted option deserves a close read. You can run the executor on your own hardware, and the docs are candid about what that means: all connections are outbound and terminate at OpenAI's environment service. Your sandbox phones home for instructions. Data residency is US-only. And a line in the docs that should get more attention than it's getting: choosing a self-hosted sandbox does not make the API eligible for zero data retention. You keep the compute. The control plane, the session state, and the logs sit on OpenAI's side of the wire.

None of this is hidden. It's stated plainly in documentation most people will skim on the way to the quickstart. That's exactly why it deserves a close read instead of a skim. The architecture encodes the business model.

Free has a track record

OpenAI has run this play before. ChatGPT launched free in November 2022, built a daily habit for hundreds of millions of people, and now sits underneath premium tiers and enterprise contracts. The free Codex CLI pulled developer workflows toward OpenAI models the same way. In each case the free layer captured the workflow, and the paid layers captured the dependencies above and below it.

The bet with the Agents API is that orchestration becomes the workflow. Once sessions, artifacts, and skills accumulate inside one vendor's session store, switching costs stop being about model quality. Your agent's memory of the last three weeks of work does not port to a competitor. Model benchmarks are portable. Session history is not.

The honest counterargument

I'd be a hypocrite to write this from a pedestal. My production agents run on a rented VPS with a cron file, a queue, and error handling I keep meaning to improve. Building durable sessions properly would take me weeks I don't have. For a solo builder, OpenAI's bundle is cheaper than my weekends, probably even at whatever general availability pricing turns out to be. Orchestration is undifferentiated heavy lifting, and handing it to a specialist is usually correct engineering.

Two things temper that. First, the zero is a beta price. The announcement commits to no fees for the public beta and says the team will iterate toward general availability, which means the real terms don't exist yet. Second, look at the partner list. Cloudflare, Modal, E2B, and the rest are building sandbox businesses on top of a platform whose owner also sells a first-party hosted sandbox. In every platform story I can think of, the owner's own option eventually gets the best shelf.

What I'm watching

The general availability pricing for the orchestration layer, because that number will reveal what the beta was for.

Whether Anthropic and Google answer with equivalent session APIs. If they do, the orchestration surface becomes the new battleground, the way context windows were in 2024.

Any movement on data residency outside the US, and on retention terms for regulated users. Enterprise adoption stalls on exactly these two lines.

Whether the nine sandbox partners keep first-class integration status or quietly drift behind the OpenAI-hosted option.

The Agents API is good engineering sold at an introductory price. I'll probably move part of my stack onto it, because it beats what I maintain myself on a weeknight. That's the free tier doing its job, and as of Wednesday it's doing it to me.