OpenAI Agents API: Computer Use, Decisions API & Pricing

👁 123 reads

OpenAI’s Agents API received computer use support at DevDay 2026: agents open browsers in OpenAI’s cloud and click through sites to complete tasks. Alongside it came Decisions API on the Luna model for rapid decisions from a closed list of responses, and Amazon is launching the same engine as Bedrock Managed Agents. We explain how it works, what it costs in dollars, and where the security pitfalls lie. Code comes from the official documentation.

What Is Agents API

OpenAI released Agents API in public beta on September 10, 2026. It’s the same infrastructure that powers Codex: OpenAI runs the agent loop, manages sessions, summarizes earlier context when approaching limits, and resumes work after breaks. Your application sends tasks and receives events. At DevDay 2026 (September 29), computer use support arrived, and the API gained Codex’s multi-agent features, tool search, and tool invocation from code. You’ll find the full list of evening announcements in our live coverage of OpenAI DevDay 2026.

The documentation describes four concepts:

• Agent - model, instructions, tools, and MCP servers.
• Environment - optional sandbox or computer: hosted by OpenAI, your own, or at a partner (including Cloudflare, E2B, Modal, Daytona, Vercel).
• Session - persistent agent instance that accepts successive tasks.
• Events and items - what you send to the agent and what it returns during work.

OpenAI’s sandbox is Linux with Python, Node.js, and command-line tools, in three sizes: small (1 vCPU, 1 GB), medium (2 vCPU, 4 GB, default), and large (4 vCPU, 16 GB). You can enable, disable, or restrict network access to a list of 1 to 100 domains. When there’s no activity or keepalive signals for an hour, the sandbox may be deleted.

First Agent: Code from Documentation

Below is an example from the official quickstart (Python), reproduced without changes. The agent creates a tree.py script in the sandbox, runs it, and shows the result. The API key needs api.agents.read, api.agents.write, and api.responses.write permissions. We did not run this example on our own account.

from openai import OpenAI

with OpenAI() as client:
    with client.beta.agents.sessions.create(
        agent={
            "model": "gpt-6-astra",
            "instructions": "Write clean code, run it, and report the actual output.",
        },
        environment={"type": "openai_hosted"},
        input="Create tree.py, a Python script that prints a readable tree of the files in the current directory. Run it and show me the output.",
        stream=True,
    ) as events:
        for event in events:
            print(event.to_json(indent=None), flush=True)

Source: Agents API quickstart. For cURL requests you must add the OpenAI-Beta: agents=v1 header; SDKs do this automatically.

Computer Use: Agent in Browser

To let an agent use a browser, the documentation instructs you to add {"type": "computer_use"} to tools and set the environment to openai_hosted with desktop enabled. The browser runs in OpenAI’s environment, not on your computer. Code from the documentation that creates such a session (without running a task):

session = client.beta.agents.sessions.create(
    agent={
        "model": "gpt-6-astra",
        "instructions": "Read public documentation in the browser. Do not sign in or change any website data. Report the page title and URL you find.",
        "tools": [{"type": "computer_use", "include_screenshots": True}],
    },
    environment={
        "type": "openai_hosted",
        "desktop": {"enabled": True},
        "network": {"access": "enabled"},
    },
)
print("Session ID:", session.id, flush=True)

Next, the application opens an event stream, sends a task (agent.session.input.message event), and handles approval requests. Each browser action is a computer_use_call item with status and optional screenshot. Screenshots are disabled by default in API responses because they may contain account data.

According to OpenAI’s summary, computer use is available through the API, and in Codex and ChatGPT Work on Pro 500 and Enterprise plans. We describe the same feature in the application in our piece on Codex in the cloud.

Security: Agent with Computer Access

The day before DevDay, OpenAI halted the GPT-6.1 Astra launch because the model in testing went beyond task scope and permissions. In our review of OpenAI’s incident reports, six of nine cases involved agents working without live supervision. That’s why it’s worth reading what the computer use documentation says about boundaries:

• Approval for every domain. The browser asks for approval at each new origin, including public ones. Enabling network access does not replace this approval.
• Domain approval is not action approval. The documentation explicitly warns that approving a domain does not require confirmation before purchases or data deletion. If you need that, restrict the browser to resources that don’t allow such actions, or use your own environment.
• Page content is untrusted. A page cannot grant the agent permissions or override user instructions.
• Passwords outside the model. Login happens via a separate event, and entered values don’t go to the model input or session history. Only the main agent can log in, not sub-agents. Passkeys and QR code login are not supported.

An example function from the documentation asks the user about each domain and defaults to deny, and cancels login requests. Our opinion: this is a good default pattern for production, because an agent that can’t visit new sites on its own has much less room for error.

How Much Does Agents API Cost

OpenAI does not charge for the Agents API itself. You pay for tokens from the chosen model, for tools according to their rates, and for sandboxes according to container pricing. We found no separate line item for computer use in the pricing. Standard rates per million tokens (context up to 272k tokens), excluding VAT:

The pricing notes that qualifying container sessions are billed per minute, with a 5-minute minimum. If speed matters, there’s also Ultrafast mode for GPT-6 Astra ($60 input and $300 output per million tokens).

Decisions API: Fast Decisions on Luna Model

Decisions API answers questions defined by the developer, choosing a response from a closed list. You provide context as text or images. The result is used for content classification, routing tickets, or choosing an agent’s next step. According to The New Stack, the API returns responses with confidence scores in 150 ms, versus 1.6 s for regular GPT-6 Luna, and The Decoder reports that OpenAI cites a tenfold speedup.

Status as of September 29: limited preview, broad availability “in the coming days.” There’s no public documentation or pricing, so we’re not showing code. We won’t guess endpoint names. This is the same model class as Jev from TypeSafe, which we covered in our piece on TypeSafe Jev and System One.

Amazon Bedrock Managed Agents

Amazon Bedrock Managed Agents, powered by OpenAI is the AWS version of Agents API. The OpenAI harness and model inference run in Amazon Bedrock, and commands and tools are executed by AgentCore Runtime or your own infrastructure. Authentication goes through IAM and SigV4 signatures, not OpenAI keys. AWS states that data does not leave AWS, and the service is in limited preview. OpenAI notes that shared concepts do not mean identical APIs, so examples from OpenAI’s documentation cannot be transferred one-to-one. Regions, limits, and pricing must be checked with AWS.

Frequently Asked Questions

How much does Agents API cost?

The API itself costs nothing. You pay for model tokens (e.g., GPT-6.1 Sol: $2 input and $10 output per million tokens), tools, and containers according to OpenAI’s pricing.

Does Agents API support Zero Data Retention?

No. The documentation states that Agents API supports data residency only in the US and is not covered by ZDR, even with your own sandbox.

Can an agent with computer use make purchases on its own?

Yes, if the site allows it. Approval covers the domain, not individual actions, which is why OpenAI advises restricting the browser or using your own environment with confirmations.

When will Decisions API be available to everyone?

OpenAI announced broad availability in the coming days. As of September 29, 2026, the API was in limited preview with no pricing.

Kinetyra