GPT-6.1 Terra: OpenAI’s Computer Use Model, Pricing and Use Cases

Last updated: 9 November 2026

What GPT-6.1 Terra is for

OpenAI’s GPT-6 generation has several models for different jobs. Terra is the one designed to do practical work quickly and cheaply: clicking through web apps, filling in forms, reading screens and documents, writing and running code, and handling everyday office tasks such as drafting proposals, reading technical drawings and editing video, according to OpenAI’s launch materials.

“Computer use” means the model looks at a screen, decides what to click or type, and repeats until the task is done. It is how agents like ChatGPT’s dots and Codex operate software that has no API. For the wider model family, see GPT-6 models explained.

Specifications

GPT-6.1 Terra pricing

Two pricing details are easy to miss:

• Long-context surcharge. Requests with more than 272,000 input tokens are charged at twice the input and cache rates and 1.5 times the output rate, and this applies to the entire request.
• Cache writes. Writing to the prompt cache costs $1.25 per million tokens for 30 minutes or $2.50 for 12 hours, while reading cached input is very cheap at $0.05.

Reasoning tokens count as output, so higher reasoning effort costs more. See prompt caching explained, batch API savings and how to estimate AI API costs.

What OpenAI says about performance

All figures below are reported by OpenAI and have not been independently verified at the time of writing.

The pattern is clear even with healthy scepticism about vendor benchmarks: Terra is much better than the previous Terra generation, and it is aimed at speed and cost rather than topping every leaderboard.

Where you can use Terra

Terra vs Sol vs Astra: which to use

Many teams route tasks between models: Terra for routine steps and a stronger model for planning or review. See model routing explained and the cheapest AI model for each task.

Practical use cases

• Marketing operations: pull figures from several web dashboards that lack APIs and assemble a weekly report.
• Ecommerce: update product listings across marketplace seller portals.
• Finance and admin: download invoices from supplier portals and log them in a spreadsheet.
• QA testing: click through a web app to check that key flows work after a release.
• Research: visit competitor sites and record prices and features in a table.
• Coding: fix failing tests and small bugs in a repository through Codex.

Cost example

Suppose an agent task uses 200,000 input tokens (mostly screenshots and page text, half of it cached) and 20,000 output tokens. At Terra’s standard rates:

• Uncached input: 100,000 tokens at $1 per million = $0.10
• Cached input: 100,000 tokens at $0.05 per million = $0.005
• Output: 20,000 tokens at $5 per million = $0.10
• Total: about $0.21 per task, before any cache write charges.

The same task at Sol’s rates would cost roughly twice as much. Real costs depend on screenshots, reasoning effort and retries, so test on your own workload.

Risks and limits of computer use

• Mistakes in real systems. A misclick can submit the wrong form. Start in test environments and require approval for consequential actions.
• Prompt injection. Web pages can contain text designed to mislead agents. See AI agent security.
• Logins and payments. Keep credentials and payments under human control.
• Site terms. Some websites prohibit automated access. Check before automating.
• Long-context cost. Watch the 272,000-token threshold.

What it means for website owners

Cheaper, faster computer use means more agents visiting websites on people’s behalf. Sites that are easy for agents to read and use, with clear text, labelled forms and plain prices, will be easier for these agents to work with. See how to make your website agent-ready.

Calling Terra through the API

A minimal request with OpenAI’s Python SDK looks like this. Check OpenAI’s documentation for the current computer use tool configuration, which needs extra setup for screenshots and actions.

from openai import OpenAI
client = OpenAI()

resp = client.responses.create(
    model="gpt-6.1-terra",
    reasoning={"effort": "low"},
    input="Extract the invoice number, date and total from this text: ...",
)
print(resp.output_text)

Start with low reasoning for extraction and simple steps, and raise it only when results fall short.

Choosing a reasoning effort level

OpenAI’s own figures show the factual error rate falling sharply at higher reasoning, from 16.7% at low to 4.3% at maximum on difficult prompts. For anything factual that customers will see, test higher levels or add a verification step.

Designing a safe computer use workflow

1. Start in a sandbox: a test account or staging environment, not production.
2. Limit permissions: give the agent access only to the apps and data the task needs.
3. Write precise instructions: name the site, the steps and what “done” looks like.
4. Require approval before submitting forms, sending messages or spending money.
5. Log every action with screenshots so you can review what happened.
6. Add a stop rule: if the page looks unexpected, the agent should stop and ask.
7. Review results regularly and refine instructions.

Terra in ChatGPT Work and Codex: tips

• Be specific about the output format, such as “a table with columns A, B and C”.
• Give examples of good results for repeated tasks.
• Break big jobs into smaller tasks so errors are caught early.
• Watch usage limits: Codex and ChatGPT Work tasks count toward your plan’s limits.

See how to use ChatGPT agent and ChatGPT message limits.

Terra compared with other computer use options

Where a site has a proper API, a fixed automation is usually cheaper and more reliable than computer use. Use computer use for interfaces without APIs.

Frequently asked questions