What OpenAI actually shipped
The pitch is simple, and it is a real shift in what you buy. Until now, "building an agent" meant building a loop: you wrote the planner, you persisted the conversation state, you handled context compaction when the window filled up, you caught failures and retried, and you wired up tool calls. OpenAI now runs that loop for you and exposes it as a service.
That is the part the marketing glosses over: session persistence, context compaction, workflow orchestration, recovery and multi-agent delegation are all managed by OpenAI. You are not buying a library. You are buying a running system, and you rent it by the token.
Access is gated behind a beta header — OpenAI-Beta: agents=v1 — which is the honest signal here. Versioned beta headers mean the surface can change under you. If you have ever shipped against a preview API, you know exactly what that costs in maintenance time. Full stop.
Four objects, and that's the mental model
The architecture is deliberately small. There are four structural concepts:
Agent, Environment, Session, Events and items
Agent is the configuration — instructions, model, tools. Environment is where the code actually executes. Session is the durable thread that survives across turns and retries. Events and items are the stream of what happened: messages, tool calls, file writes, errors.
Compare that with the hand-rolled version most teams run today, where "session" is a row in Postgres, "environment" is a Docker container you babysit, and "events" is a log file nobody reads until something breaks. The API is basically a productised version of that stack — which is precisely why it will be attractive to small teams and mildly annoying to platform engineers who already built it.
What it costs — and what it doesn't
This is the part where OpenAI's pricing is genuinely straightforward: there is no additional platform fee for the Agents API. You pay standard model token rates for whatever model the agent uses (the example given is gpt-6-astra) plus standard container rates for OpenAI-hosted sandboxes.
Set that against the rest of the market and the structure looks competitive. xAI's Grok 4.6, for instance, is priced at $2.00 per 1M input tokens, $0.50 per 1M cached input tokens and $6.00 per 1M output tokens for prompts under 200k tokens — a raw model price with nothing around it. OpenAI's offer is the opposite bet: the same kind of per-token billing, but the orchestration layer comes bundled. Whether that is cheaper depends entirely on how much engineering time you were spending on retries, compaction and session state. For a two-person startup, that number is large. For a platform team with a working harness, it is close to zero.
One practical note for European buyers: OpenAI bills in US dollars, and EU business customers should expect VAT on top of the listed token rates. Budget the euro equivalent with that in mind rather than comparing headline numbers straight across.
Nine sandboxes, your VPC, or OpenAI's own
Execution is where the API gets flexible. You can run the agent in OpenAI-hosted containerised sandboxes, inside your own VPC or infrastructure, or on a partner cloud sandbox. Nine external providers were integrated at launch: Blaxel AI, Cloudflare Dev, Daytona, DigitalOcean, E2B, Modal, Oracle Cloud, Runloop AI and Vercel. There is also support for Model Context Protocol (MCP) servers, which matters more than it sounds — MCP is becoming the common way to hand an agent tools without writing bespoke glue for every integration.
That is a healthy list, and several of those providers operate European regions. But "the sandbox runs in Frankfurt" is not the same claim as "your agent data stays in Frankfurt", and the beta documentation does not support the second claim.
The European problem: US-only residency, no Zero Data Retention
Here is the limit that will decide adoption in EU enterprises. The public beta ships with US-only data residency and no support for Zero Data Retention. Both are stated limitations, not rumours.
For a European company, that combination is a conversation with legal, not with engineering. Sessions are persistent by design — that is the feature. Persistence means prompts, tool outputs, file contents and intermediate reasoning can sit on infrastructure outside the EEA, and without ZDR you cannot promise a client that nothing is retained after the call. If you process personal data, or anything under professional secrecy, that is a transfer question under GDPR, and it needs a transfer impact assessment rather than a shrug and standard contractual clauses.
Choosing a sandbox in your own VPC does not automatically solve it either. The sandbox is the execution layer; session state and orchestration are the service. Until OpenAI confirms an EU residency option for the service layer, treat "bring your own compute" as a performance and control feature, not a compliance fix. Ask the question in writing before you architect around it.
The regulatory clock has already started
Timing matters because the EU's enforcement posture changed this summer. Since August 2, 2026, the European Commission and the EU AI Office hold direct enforcement powers over general-purpose AI models — binding requests for model access, information demands, forced recalls and financial penalties. The old grace period of voluntary adherence is gone.
At the same time, Article 50 transparency obligations are mandatory: users must be told when they are interacting with an AI system, deepfakes must be visibly labelled, and synthetic content must carry machine-readable marking. If you ship an agent that talks to end users or generates media, that obligation lands on you as the deployer — not on OpenAI. The complicated bit for agent builders is that Article 50 was written with outputs in mind, not with autonomous multi-step tool use in mind, so the honest answer is that you need your own disclosure layer in the product UI.
There is one piece of genuine relief: under the EU's Digital Omnibus simplification package, the compliance deadline for standalone high-risk systems listed in Annex III was deferred to December 2, 2027. That buys planning time. It does not remove the obligation.
Who should actually try this now
If you are building coding agents, internal automation or document pipelines, and your data is not sensitive personal data, the Agents API removes a serious amount of plumbing. The token-plus-container billing model means you can estimate cost per task the same way you estimate model cost today, and the four-object model is small enough to learn in an afternoon.
If you are building for EU regulated clients, the calculation is different. The orchestration layer is excellent, but residency is the blocker, and no amount of partner sandbox choice fixes it in the beta. Start with a proof of concept on non-sensitive data, keep your session abstraction replaceable — because you may need to swap the harness out — and watch for an EU residency announcement. If you want to understand what the self-hosted alternative costs in practice, our AI Arena benchmark rig is where we measure that trade-off on real hardware. More background on the European AI market is in the magazine.
Does running the sandbox in my own VPC keep my agent data in the EU?
Not necessarily. The sandbox is only the execution layer. Session persistence, orchestration and recovery are handled by OpenAI's service, and the beta is documented as US-only for data residency with no Zero Data Retention support. If residency is a hard requirement, get written confirmation about the service layer before assuming your own VPC solves it.
Is there an EU-hosted equivalent I can use today?
There is no drop-in managed equivalent of this harness offered by a European provider right now. The realistic alternatives are running your own orchestration code on EU infrastructure, or using a European platform's API and accepting less managed functionality. Mistral's La Plateforme, for instance, remains the most obvious EU-based option for API work, but the scope of what it manages differs.
Do I have to disclose that a user is talking to an agent I built with this API?
Yes, if you deploy in the EU. Article 50 transparency obligations have been mandatory since August 2, 2026, and they require telling users when they interact with an AI system. The obligation sits with you as the deployer, not with OpenAI, so plan a disclosure layer in your own interface.
