Computer use or API? When a screen agent is the right call

OpenAI released GPT-6.1 Terra on 26 October 2026. It's a model built for computer use, priced at $1 per million input tokens and $5 per million output tokens. OpenAI reports it is about 9x cheaper per task than Claude Opus 5.5 on its own computer-use benchmark.

Artificial Analysis ranks it independently on general tasks, not computer use: 48 on its Intelligence Index, 25th of 221 models, for $0.40 per task at max effort. Its speed isn't measured yet.

Computer use means the agent works through the screen: it takes a screenshot, decides where to click or what to type, acts, then takes another screenshot. It works on any interface a person can use. Cheaper doesn't mean it should replace your APIs.

When an API or an MCP server covers the action, use it

• —Errors: a typed call either succeeds or returns an error. A click can land on a popup or a button that moved, without any error.
• —Cost: one call costs a few hundred tokens. A screenshot loop costs thousands per step.
• —Permissions: you can scope an IAM role or an OAuth token to one action and log it. A screen agent has the rights of the whole session.
• —Injection: anything displayed on screen, an email or a web page, can carry instructions for the agent.

When computer use makes sense

• —There is no API: legacy desktop software, supplier portals, internal tools without an SDK.
• —The task is visual: reading a technical drawing, editing a video, checking that a page renders correctly.
• —It's a one-off job, and writing the integration would cost more than doing the task once.

If a screen workflow runs every week, it's a missing integration: keep computer use for now and build the API next. OpenAI's own example uses both. Reading a scanned invoice needs vision. Entering it into accounting software should go through the software's API when one exists.

Where the browser runs

The model doesn't drive a browser on its own: it returns actions, and something has to run them.

• —Hosted by the product: ChatGPT and Codex run the browser on the vendor's side.
• —An API tool and your own executor: the model returns actions, your code runs them, for example Playwright in a container.
• —A local Playwright MCP server, one per user, using their own browser and logged-in sessions.
• —A managed remote browser: on AWS, AgentCore Browser runs isolated sessions of up to 8 hours, with live view, CloudTrail logging and session recording to S3.

For a team of users, a managed remote browser is easier than installing an MCP server on every machine. What you give up is the users' logged-in sessions, so authentication has to be handled separately.

Sources

• —Introducing GPT-6.1 Terra, OpenAI
• —GPT-6.1 Terra, Artificial Analysis
• —Amazon Bedrock AgentCore Browser, AWS documentation