OpenAI’s Agents API hands developers the same engine that keeps its own agents running for hours. Here’s what actually changed, what it costs, and where the beta still bites.
Picture an automation that has to run for six hours without a person watching it. It pulls a week of invoices, checks each one against a ledger, flags the mismatches, and writes a plain summary before anyone reaches their desk. Until recently, building that meant writing your own scaffolding to keep the task alive, feed it fresh context as it grew, restart it when a single step failed, and stop it from quietly running your cloud bill into the ground.
On September 10, 2026, OpenAI put most of that plumbing behind one API call. The release, the Agents API, is now in public beta, and it changes what a cloud agent can realistically do without an engineer babysitting every run.
The Short Version
The Agents API is a managed service that runs long-lived AI agents in the cloud for you. You describe the task, pick a model, list the tools it can use, and choose where its code executes. OpenAI keeps the session alive, trims the context as it grows, recovers from failures, and coordinates any subagents the job spins up. There is no separate fee for the harness itself, so you pay for the tokens and tools each agent burns. The catch is honest and worth reading twice: it is still beta, US-only for now, and does not yet support zero data retention.
What OpenAI actually shipped
The Agents API exposes the Codex harness, the same orchestration engine that already runs long tasks inside Codex and enterprise ChatGPT, and puts it behind a single call. In OpenAI’s own framing, you can spin up a production-ready agent by naming the task, the model, the tools, and the environment, and the service handles the rest. You can read the full write-up in OpenAI’s announcement.
That “rest” is the interesting part. The harness is versioned and gets updated alongside new model releases, so the agent scaffolding improves when the models do, rather than aging in your codebase. According to MarkTechPost’s breakdown, the API is organized around four pieces: an Agent (the model plus its instructions and tools), a Session (a durable task that survives across many turns), a Sandbox (where any code actually runs), and the tools and events flowing in and out.
OpenAI also pointed to early customer numbers. A firm called SafetyKit reported a 60% cost cut per case after moving a workflow onto the harness, and Hypha said it saw 86% fewer failed agent responses. Those are company-supplied figures, not independent benchmarks, so treat them as direction rather than proof.
Why a plain model API wasn’t enough
A normal model API gives you one answer and forgets you exist. Ask a question, get a response, done. That works fine for a chatbot reply or a quick summary.
An agent that works for hours needs much more than a single reply. It has to hold context as the session grows past the model’s memory, connect and orchestrate tools in the right order, run code somewhere safe, save files, and pick itself back up when a step throws an error at 3 a.m. Developers used to write all of that themselves, and honestly, most of the pain in shipping an agent lived in that glue code, not in the prompts.
| Plain model API | Agents API | |
|---|---|---|
| What you get back | One response, then it forgets you | A durable session that can run for hours or days |
| Session state | You track and rebuild it yourself | Managed, with context compaction and recovery |
| Tools and code | You orchestrate and sandbox them | Orchestrated for you; code runs in a sandbox you pick |
| What you maintain | The whole agent loop | Just the tools and the task design |
The Agents API turns that loop into a managed service. Context compaction, recovery, and subagent coordination now happen on OpenAI’s side. You still supply the tools and decide where code runs, but you are no longer maintaining a small distributed system just to keep one long task from falling over.
Where your agent actually runs
One decision the launch forces early is where the agent’s code executes. This is the sandbox, and OpenAI gives you three routes:
- OpenAI-hosted: a zero-setup isolated environment. Fastest to start, and you pay for the container time on top of tokens.
- Your own VPC: the agent runs a small server inside your infrastructure over an outbound-only connection, so code and data never leave your cloud perimeter. OpenAI still manages the session layer.
- Partner sandboxes: nine outside providers, including Cloudflare, Vercel, DigitalOcean, Oracle, Modal, and E2B, each with its own compute tiers.
The partner route is where some clever billing shows up. Cloudflare, for one, snapshots and sleeps its container after about 30 seconds of no activity and thaws it in milliseconds on the next task, so an idle agent stops costing money between jobs. For automations that fire in bursts rather than running flat out, that idle behavior can matter more to the monthly bill than the model choice.
What it costs
Here is the part people scrolled down for. There is no separate fee for the Agents API or the harness. You pay standard rates for the model tokens and tool calls your agent uses, plus container time if you pick the OpenAI-hosted sandbox. Choose your own VPC or a partner sandbox, and the compute bill goes to that provider instead.
The pricing is refreshingly simple, but simple is not the same as cheap. A long-running agent that keeps a session warm and calls tools repeatedly can quietly rack up tokens across a multi-hour job. The cost lives in how long the agent runs and how chatty it is, not in a line item labeled “agents.”
The catches the launch posts skipped
Most of the coverage repeated the good news. The parts worth slowing down for are the limits, because they decide whether you can actually put this in production today.
Three stand out. First, it is a public beta, so there is no general-availability date and the behavior can shift. Second, data residency is US-only right now, which rules out teams with EU or other regional requirements. Third, and this is the big one for regulated work, zero data retention is not supported yet, even on self-hosted sandboxes. If your compliance team requires ZDR, this is a hard stop until it lands.
There is a softer catch too, and developers said it out loud in the first week. Shipping an API is not the same as getting a reliable agent through deploy. One engineer put it bluntly: OpenAI shipped an Agents API, and they still could not get a single agent all the way to production. The managed harness removes a real chunk of the plumbing, but the hard parts of agent design, deciding what the agent should and should not do, and testing it against messy real inputs, are still yours to solve.
Worth noting: OpenAI is not alone here. Anthropic and the big cloud providers now offer their own managed agent runtimes, so this is a competitive market forming quickly, not a solo launch. That is good for prices and portability, and it means locking into one provider’s harness is a choice to make on purpose. For more on how these tools reshape business operations, browse SciExaminer’s Business coverage.
What To Know
- The Agents API runs long-lived cloud agents for you, using the same Codex harness behind OpenAI’s own agents.
- It handles sessions, context compaction, failure recovery, and subagents, so you stop writing that glue code yourself.
- You choose the sandbox: OpenAI-hosted, your own VPC, or one of nine partner providers.
- No harness fee. You pay for tokens, tools, and container time, and long runs can add up.
- It is beta, US-only, and has no zero data retention yet, which blocks some regulated and EU workloads.
Frequently Asked Questions
What is the OpenAI Agents API in plain terms?
The OpenAI Agents API is a managed service that runs AI agents in the cloud over long tasks. You define the agent, and OpenAI keeps the session alive, manages memory, and recovers from failures instead of making you build that infrastructure.
How is it different from just calling a model?
A model API returns one response and keeps no state. The Agents API maintains a durable session that can run for hours or days, orchestrates tools, runs code in a sandbox, and picks up after errors on its own.
How much does the OpenAI Agents API cost?
There is no separate fee for the API or the harness. You pay standard rates for model tokens and tool usage, plus container time if you use the OpenAI-hosted sandbox. Your own VPC or a partner sandbox shifts the compute bill to that provider.
Can I keep the agent’s code inside my own cloud?
Yes. Alongside the OpenAI-hosted option, you can run the sandbox in your own VPC over an outbound-only connection so code stays in your perimeter, or use one of nine partner sandboxes such as Cloudflare or Vercel.
Is it ready for production?
It is a public beta with no general-availability date, US-only data residency, and no zero data retention support yet. It removes a lot of orchestration work, but teams with EU or strict compliance needs should wait.
What This Means
The Agents API is a plumbing upgrade, not a headline model, and plumbing upgrades are what quietly move a technology from demos into daily work. By handing developers the harness that already runs its own long tasks, OpenAI lowers the bar for the kind of six-hour, run-while-you-sleep automation that used to demand a dedicated team. If you build with AI, the smart move this month is a small pilot on the hosted sandbox to feel out cost and reliability, while keeping the beta limits in plain view. Keep an eye on the Technology section for how the managed-agent race develops from here.