How much does the OpenAI Agents API cost?

There is no separate fee for the Agents API as of September 2026. You pay the selected model's normal token rates (for example GPT-6 Astra at $10 per million input and $50 per million output tokens), OpenAI's built-in tools at their standard rates, and — if you use an OpenAI-hosted sandbox — the standard container rates listed on the pricing page. Self-hosted environments and partner sandboxes (Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, Vercel) are billed by those providers.

What are the limits of the Agents API in beta?

Three worth knowing before you build on it: the API is in public beta with no GA date; it supports data residency only in the United States; and it does not support Zero Data Retention — choosing a self-hosted sandbox does not make it ZDR-eligible. Sessions retain state on OpenAI's side so you can continue work across turns; you can delete sessions and published artifacts when you no longer need them. Sessions are created under the beta endpoint POST /v1/agents/sessions with the OpenAI-Beta: agents=v1 header.

Is the Agents API the same as the OpenAI Agents SDK or the Responses API?

No. OpenAI's own docs state that an Agents API session, an Agents SDK session, a Responses conversation and a sandbox are different resources. The Responses API is a stateless model call you orchestrate yourself; the Agents SDK is an open-source library that runs the loop in your process; the Agents API is a hosted loop that OpenAI runs for you, with a sandbox attached. Pick the SDK when you need full control in your own code, the Agents API when you want long-running cloud agents without maintaining a harness.

Quick Answer

What Is the OpenAI Agents API? Codex Harness Explained

Q: What is the OpenAI Agents API?

The Agents API is an OpenAI-managed service, released in public beta on September 10, 2026, that gives developers the same agent harness and infrastructure that runs Codex and ChatGPT Work. You create a session with one API call — model, instructions, tools, MCP servers and an environment — and OpenAI runs the orchestration loop: executing commands in a sandbox, compacting context as a session approaches its limit, delegating to parallel subagents, steering mid-turn and resuming sessions later. The harness itself is the open-source Codex codebase on GitHub; OpenAI operates and maintains the hosted version.

Published: September 11, 2026

The Short Answer

On September 10, 2026, OpenAI released the Agents API in public beta: a hosted version of the Codex harness that runs long-lived agent sessions for you. You describe the agent (model, instructions, tools, MCP servers), choose where it runs (OpenAI-hosted sandbox, your own infrastructure, or a partner sandbox), and send it tasks. OpenAI handles orchestration, context compaction, subagent delegation, steering and recovery. There is no surcharge — you pay token, tool and container rates.

	Detail (as of September 11, 2026)
Status	Public beta, all developers, no GA date
Endpoint	`POST /v1/agents/sessions` with header `OpenAI-Beta: agents=v1`; SDK call `client.beta.agents.sessions.create(...)`
Harness	Same open-source Codex harness that powers Codex and ChatGPT Work (github.com/openai/codex)
Environments	OpenAI-hosted sandbox · self-hosted (`type: "self_hosted"`) · partner sandboxes: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, Vercel
Built in	Context compaction, tool search, programmatic tool calling, multi-agent (`max_concurrent_subagents`), MCP, web search, skills, artifacts, webhooks/streaming
Pricing	Model tokens at API rates + built-in tools at standard rates + OpenAI sandboxes at standard container rates. No Agents API fee.
Data controls	US data residency only; no Zero Data Retention, even with a self-hosted sandbox

What you actually get

OpenAI’s framing: after scaling Codex and ChatGPT Work to millions of users, the hard part of agents turned out to be the harness — context management, efficient tool use, subagent coordination — plus infrastructure that keeps an agent running for days with files, code execution and intermediate results. The Agents API packages both.

The four core concepts in the docs:

Agent — the model, instructions, tools and MCP servers.
Environment — an optional sandbox or computer where the agent reads files, loads skills and runs commands.
Session — a durable instance of the agent that works on tasks and accepts more input.
Events and items — what you send in and what comes out, via streaming or webhooks.

A session’s lifecycle is: create (OpenAI provisions the environment) → give it a task → follow progress by stream or webhook → send another task to the same session, or steer it mid-turn.

The harness features that matter

Compaction. As a session approaches the model’s context limit, the API automatically summarises earlier context while preserving what the agent needs to continue. Workflows can span multiple context windows without you writing compaction logic.

Tool search + programmatic tool calling. Tool definitions are loaded on demand to save tokens and preserve the prompt cache. Once tools are available, the agent can call them in parallel, chain them and filter results in code, bringing only the relevant slice back into context — the same trick that makes Codex efficient on large datasets.

Subagents. With multi_agent: { enabled: true, max_concurrent_subagents: N }, the main agent splits work and delegates to subagents that each keep their own context, then merges results. OpenAI’s example is comparing release notes in parallel.

Steering and resuming. You can redirect an agent during a turn and pick a session back up where it left off.

Where the agent runs

This is the design decision OpenAI made differently from most competitors: the harness is hosted, but the compute is your choice.

OpenAI-hosted sandbox — the same sandboxing infrastructure behind Codex and ChatGPT, provisioned and managed by OpenAI, configurable with your files, packages, skills and plugins. Billed at container rates.
Self-hosted — the agent’s commands run on your infrastructure (workspace_directory, capability_directories), inside your VPC and network policy.
Partner sandboxes — first-class integrations with Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel, covering managed environments, VPC deployments, specific secret/file storage, and different CPU/GPU/memory and cold-start profiles.

Note the fine print: self-hosting the sandbox keeps your files home, but the session state, transcripts and orchestration still live with OpenAI in the US, and the endpoint is not ZDR-eligible. Regulated teams that need ZDR should stay on the Responses API with their own loop for now.

How it compares (September 2026)

	OpenAI Agents API	OpenAI Agents SDK	Anthropic Managed Agents / Claude Agent SDK
Who runs the loop	OpenAI (hosted Codex harness)	You, in-process (open-source library)	Anthropic (Managed Agents) or you (Agent SDK)
Sandbox	OpenAI-hosted, self-hosted or 9 partners	Whatever you wire up	Anthropic-managed cloud sandbox or self-hosted sandbox (GA on Claude API since mid-2026)
Extra fee	None — tokens + tools + containers	None	Managed Agents bill session-hours plus tokens
ZDR	No	Depends on the model endpoint you call	Depends on plan/endpoint
Harness source	Open (github.com/openai/codex)	Open	Agent SDK open; managed control plane closed

The closest analogue is Anthropic’s Managed Agents plus the Claude Agent SDK: both vendors now sell “our production harness, your tools, your choice of sandbox.” The difference in September 2026 is pricing shape — OpenAI charges nothing for the harness itself, Anthropic meters session-hours — and OpenAI’s explicit ZDR gap.

Who should use it

Yes: teams building long-running cloud agents (incident responders, GitHub issue investigators, data analysts over a warehouse, document reviewers — all of which OpenAI ships as showcase apps) who do not want to maintain compaction, tool routing and subagent orchestration themselves.
Yes, carefully: anyone already on the Agents SDK who wants to move heavy sessions off their own servers; sessions, SDK sessions and Responses conversations are separate resources, so this is a migration, not a flag flip.
Not yet: ZDR-mandated workloads, non-US data residency, or products that need GA-level stability guarantees.