OpenAI's Agents API Is in Public Beta: 9 Sandbox Partners, No New Fees, and a Data Residency Catch
OpenAI turned its internal Codex harness into a managed product on September 10. It costs nothing extra beyond token usage, runs on 9 different sandbox partners, and comes with a data residency catch most quick takes leave out.
Contents5
On September 10, OpenAI opened its Agents API to public beta, and the pitch is straightforward: the same harness and infrastructure that runs Codex is now available behind one API call, for any long-running agent, not just coding ones. If you've been hand-rolling context management, tool orchestration, and subagent coordination for your own agent product, this is OpenAI offering to do that plumbing for you.
What the Agents API actually replaces
The API organizes around four concepts: an agent (model, instructions, tools, and MCP servers), an environment (an optional sandbox for file access and command execution), a session (a durable instance of an agent working on a task), and events and items (the inputs and outputs moving through it).
That's a different layer than what OpenAI already shipped. The Agents SDK runs inside your own application and you manage the infrastructure. The Responses API hands you raw model calls and leaves orchestration entirely up to you. The Agents API is the managed version: OpenAI runs the harness, you send it a task.
Two features do most of the actual work. Automatic compaction rewrites earlier context as a session approaches its context limit, so long-running agents don't just fall over when they run out of room. Programmatic tool calling lets tools run in parallel or get chained together instead of one at a time. There's also real subagent support: a primary agent can hand a self-contained piece of work, like an independent code review or a parallel data check, to a subagent that gets its own separate context.
Nine ways to run the sandbox
Every agent that needs to touch files or run commands needs an environment, and OpenAI gives you three ways to host it.
| Option | What it means |
|---|---|
| OpenAI-hosted | Runs on the same Codex/ChatGPT infrastructure OpenAI already operates |
| Self-hosted | You run `codex exec-server` yourself and connect over WebSocket |
| Partner-hosted | Nine integrations: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel |
Nine partner options is a lot of choice, and it's a tell that OpenAI is betting on this API getting used for all kinds of agent workloads beyond coding, not just plugged into existing OpenAI-hosted infrastructure. If you're already running workloads on Cloudflare Workers or Vercel, that's presumably the point: your sandbox lives where your other infrastructure already does.
Pricing: no new fee, but the model choice matters a lot
The API itself doesn't add a separate charge. You pay standard rates for whatever model and tools you use, plus container time for hosted sandboxes. But which model you pick changes your bill dramatically, because OpenAI's current model line spans a wide range:
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| gpt-6-astra | $10 | $50 |
| gpt-5.6-terra | $2 | $12 |
| gpt-5.6-sol (promo pricing through at least Nov 21, 2026) | $4 | $20 |
| gpt-6-astra Fast Mode | $20 | $100 |
| luna | $0.20 | $1.20 |
Astra, the flagship, costs 2.5 times what Sol currently costs on its promotional rate, and 5 times what Terra costs. For an agent that runs constantly in the background rather than answering a handful of user questions a day, that multiplier compounds fast. Terra or Luna are the more realistic default for a lot of Agents API workloads unless you specifically need Astra's reasoning quality.
The catches most quick takes skip
A few limits are easy to miss if you only read the announcement post. Data residency is US-only right now, which rules this out for teams that need to keep agent data in the EU or elsewhere for compliance reasons. Zero Data Retention isn't supported either, which is the same problem from a different angle: if your industry requires ZDR, the managed Agents API isn't an option yet, and you're back to self-hosting or the Agents SDK.
Automatic context compaction also doesn't mean you're off the hook for state management. Persistent application state, the stuff that needs to survive a restart or move between sessions, is still your job to build. The managed harness handles context inside a session; it doesn't replace your database.
And if you already have a working custom orchestrator, migrating to this isn't automatically a win. How much value you get depends entirely on how much of your own infrastructure the managed harness can actually replace. Teams starting a new long-running agent project from scratch get more out of this than teams ripping out something that already works.
Who should actually use this
If you're building a new agent product and don't want to reinvent context management, tool orchestration, and subagent handoffs from scratch, the Agents API is a real shortcut, and the 9 sandbox partner options mean you're not forced onto OpenAI's own infrastructure to use it. If you're outside the US and need data residency guarantees, or you're in a regulated industry that requires Zero Data Retention, it's not ready for you yet. And if you've already built your own orchestration layer that works, read the fine print before you assume switching saves you anything.
For comparison, [DeepSeek moved the opposite direction on pricing this year](https://questloops.com/blog/deepseek-just-raised-api-prices-by-up-to-1-100-what-changed-and-what-to-do-about-it), and [Gemini's 3.7 Flash launch pricing has its own expiration date to watch](https://questloops.com/blog/gemini-3-7-flash-pricing-the-half-price-launch-has-an-expiration-date). Worth checking both against Terra and Luna before you commit token spend to one provider's Agents API.