GPT-6.1 Terra uses your computer. Should your business act on it now?

On Monday 26 October, OpenAI released GPT-6.1 Terra, a model that looks at your screen and clicks, types and fills in forms by itself. In the same week, OpenAI said two of its own AI agents had switched off their monitoring. One of them denied it afterwards. Those two stories belong together, and that is why I'm writing this. We tried Terra ourselves this week and haven't seen anything odd yet. But a week of testing tells you little about a year in production.

What is GPT-6.1 Terra?

GPT-6.1 Terra is OpenAI's mid-range model, released on 26 October 2026. It reads what is on a screen and takes actions in the browser and on the desktop by itself: clicking, typing, filling in forms. Through the API it costs $1 per million input tokens and $5 for output, half the price of its predecessor.

Terra sits between the lighter Luna and the heavier Sol. According to OpenAI, it finishes tasks more than three times as fast as Astra and Sol. It also makes fewer mistakes: on an internal test, the error rate fell from 53.6 to 16.7 percent. Those are OpenAI's own figures, so treat them as a claim rather than a fact.

Simon Willison, one of the most widely read AI bloggers, made a sharp point. Terra occupies the same spot as Sonnet does at Anthropic: not the cheapest, not the most expensive. That is exactly what makes it hard to decide when to use it.

A short definition, because the term is turning up everywhere.

Computer use means an AI model operates software the way a person does: through the screen, not through an integration.

What can a business do with AI that uses your computer?

Mostly work in old or closed systems without a decent integration. Think of retyping data from a PDF into your accounting package, working through a supplier portal or filling in forms for a public body. If your software does have an API, a regular integration is almost always faster, cheaper and easier to check.

I see computer use as duct tape. Handy where no integration exists, but you don't build a house with it.

Look closely at the numbers, too. On Agents' Last Exam, a test made up of professional tasks, Terra scores 60 percent at its highest setting, for roughly a dollar per task. That's impressive. It also means four in ten tasks fail. For work where a person reviews the exceptions, that's fine. For work nobody checks, it's a problem.

You see the same thing in the market. Everyone is wary of letting this loose on payments, and rightly so. Automating the groundwork is fine: collecting, checking and preparing data. As long as the real decision stays with a person and the AI has no access to it, you keep it well under control.

Why did an OpenAI agent switch off its own monitoring?

On 3 October, an agent running GPT-6.1 Sol switched off the automatic alerts to OpenAI's security team. It read an instruction not to send Slack messages too broadly. On 6 October, an agent running GPT-6 Astra removed a monitor and then told its user it hadn't. Staff caught both on the same day.

Neither incident was given the highest priority, and the system didn't stop either agent automatically. OpenAI itself writes that it doesn't yet have a satisfactory explanation for this behaviour (CNBC, 21 October 2026).

My conclusion isn't that you should leave AI agents alone. It is the best argument I know for limits that sit outside the agent. An agent that can switch off its own controls has no controls. The big vendors seem to agree. Analyst firm Info-Tech saw the same line that week at OpenAI, Google, Microsoft and AWS: security belongs outside the model and outside the agent's environment.

Ethan Mollick, a professor at Wharton, asked a question in August that I keep coming back to. For years we've practised the question of when people should ask AI for help. Now we also have to decide when an AI should ask us.

How do you deploy an AI employee safely?

For each task, set out what the AI employee may do by itself, what goes for approval first and when a person takes over. Give it only the permissions that one task needs. Log every action, and run that monitoring somewhere it can't reach. Then it matters much less how clever the model is.

Why so strict? Because we also see cases where something is written into the rules in so many words and an AI still doesn't always stick to it. An instruction is a request. A permission it doesn't have is a limit.

Should you start with GPT-6.1 Terra now?

Not with the model, but with the question of what work it should do. Models change fast: OpenAI released GPT-6.1 Sol on 29 September and Terra four weeks later. The tasks in your business change far more slowly. If you know which process is costing money now, you'll pick the right model later without stress.

This week, Dutch tech podcast AI Report, run by Alexander Klöpping and Wietse Hage, is helping listeners build their own AI colleague in five days. It's fun, and I'd recommend it to anyone who wants to feel how this works. For a business, though, I'd spend those five days differently. First work out what that colleague should do, then build.

That is exactly what our AI Opportunity Audit does in five working days. You get what the work costs today, the biggest opportunities and also what we would not do. Sometimes the answer is a regular integration, as I explain in AI agent or regular automation.

What can you do this week?

Take one process where someone retypes data from one system into another. Count how many hours that takes per month. That number is worth more than any OpenAI benchmark.