GPT-6.1 Terra: computer use you can finally leave running all day
OpenAI just launched GPT-6.1 Terra, a fast model that sees your screen and works across apps. It's 15 points better on agent tasks and at about a third of the cost. Here are the numbers and what they open up.
Computer-use agents just got cheap enough to leave on.
Today OpenAI released GPT-6.1 Terra. It reads what's on the screen, clicks and types across apps, and writes like a person. It also costs a fraction of what this level of capability cost yesterday. Until now, agents that could drive a screen were mostly a demo. Terra looks like the version you actually put to work.
What is GPT-6.1 Terra?
GPT-6.1 Terra is OpenAI's new fast, efficient model for visual reasoning, precise computer use and clear, natural writing. OpenAI calls it "a significant step up" from GPT-5.6 Terra, built for the screen-based tasks that fill a workday: filling out forms, reading technical drawings, drafting proposals, even editing videos.
The example in the announcement says a lot: reading incoming invoices and typing their details into accounting software. That's boring, repetitive work, and it's exactly where an agent like this pays off.
How much better is it?
OpenAI tested it on Agents' Last Exam, a benchmark of real professional tasks done on a computer. These are the results at medium reasoning effort:
- +15 percentage points over GPT-5.6 Terra
- at about one-third of the cost
- on par with GPT-6 Astra, at one-tenth of its price per token
- over 3x faster than GPT-6 Astra and GPT-6.1 Sol at completing tasks
So it's better and cheaper, and much faster on top. Usually you get to pick only one of those.
How much does GPT-6.1 Terra cost?
Standard API pricing, per million tokens:
| Price | |
|---|---|
| Input | $1.00 |
| Cached input | $0.05 |
| Output | $5.00 |
That's 50% lower than GPT-5.6 Terra and GPT-6.1 Sol. Batch and Flex are another 50% off.
Look at that cached-input price. Agents resend the same instructions and context again and again, and at five cents per million cached tokens, that repeated part costs almost nothing. That's what makes the "leave it running all day" idea realistic.
Where can you use it?
- ChatGPT Work and Codex, starting today, for Plus, Pro, Enterprise and Edu users
- The OpenAI API, as
gpt-6.1-terra - A context window of over 1 million tokens
What this opens up
For a long time the question about computer-use agents was "can it do the task?" Now it's "which tasks do I hand over first?"
Think of the work that lives between apps with no API: copying data from a PDF into a web form, checking a dashboard and logging what changed, going through a backlog of screenshots. All of that just became cheap enough to automate without doing the math twice.
I already have one in mind for myself: a nightly agent that clicks through my own sites like a real visitor and tells me what feels broken. At these prices, that's an easy experiment.
Your turn
The model is out, the price is low and the API is open. So the interesting part isn't the benchmark anymore. It's what people build with it.
What's the first workflow you'd hand to an agent that can see your screen? Tell me on X at @SergioNavarroAI.