Prince Mario-Max Schaumburg-Lippe: GPT-6.1 Terra Puts AI to Work on Your Screen

The most useful AI release of the week does not arrive with a bigger chat window. It arrives with a pair of hands. On Monday, October 26, 2026, OpenAI launched GPT-6.1 Terra, a model built to look at a computer screen, understand what it is seeing, and then click, type and move between applications until a job is done. Chief executive Sam Altman announced it on X with a short post that said plenty: “6.1 Terra is here! It is good at a lot of things, but we are especially excited for this new era of computer use.”

That sentence is the whole story, really. Terra is the second model in the GPT-6.1 series, joining GPT-6.1 Sol, and it succeeds GPT-5.6 Terra. OpenAI is pitching it less as a conversationalist and more as a careful colleague for the routine screen work that eats an office day: drafting proposals, reading technical drawings, filling in long forms, editing video, and moving invoice details into accounting software. Anyone who has copied numbers from one window into another for an hour knows exactly which hour this model wants back.

Built to act, priced to spread

Speed and price do the persuading here. OpenAI says Terra completes tasks more than three times faster than its larger siblings, GPT-6 Astra and GPT-6.1 Sol. Through the API, where developers call it gpt-6.1-terra, standard pricing is $1 per million input tokens, 5 cents per million cached input tokens and $5 per million output tokens. OpenAI says those rates are 50 percent lower than both GPT-5.6 Terra and GPT-6.1 Sol. Half price matters more than it sounds. A tool that handles one form is a demo. A tool priced so a firm can run it across thousands of forms a day is a staffing decision.

Access is broad from day one. Terra is rolling out to Plus, Pro, Enterprise and Edu users inside ChatGPT Work and Codex, with Business plans listed in the release notes, and developers can reach it through the API. Amazon Web Services made it available on its Bedrock platform the same day, a sign OpenAI expects companies to test it inside systems they already run. The model takes in text and images, carries a context window of about 1.05 million tokens with up to 128,000 output tokens, and has a knowledge cutoff of June 1, 2026, according to OpenAI’s developer documentation. In plain terms, it can hold a very long file, a long thread and a long set of instructions in mind at once.

It joins a busy week for practical AI. Readers following our coverage of Microsoft Decision 1 scoring choices in 85 milliseconds and of iFlytek Spark X2.6 putting AI on device and in the cloud will spot the same pattern: the race has moved from answering cleverly to finishing work reliably, wherever the work happens to live.

What the tests say, and who is saying it

A note of care before the numbers. Every figure in this section comes from OpenAI’s own announcement and testing. None of it has been independently verified yet, and the article treats it that way: as a strong claim from the maker, awaiting outside checks.

On Agents’ Last Exam, a test of how well an agent uses software at work, OpenAI says Terra scored 15 percentage points higher than GPT-5.6 Terra at medium reasoning effort, at about one-third the cost per million tokens. At maximum effort, the company says Terra performed on a par with GPT-6 Astra at one-tenth the price per token. On OSWorld 2.1, which asks an agent to use desktop apps for multi-step jobs such as building a 3D model from a technical drawing, OpenAI reports 53.8 percent at low reasoning effort, above what its predecessor managed at maximum effort and higher cost.

Fewer slips, steadier hands

The error figures may matter most to anyone planning to trust an agent with real files. On an internal set of prompts designed to trigger mistakes, OpenAI says Terra’s error rate at low reasoning effort fell to 16.7 percent, down from 53.6 percent for GPT-5.6 Terra, and reached 4.3 percent at maximum effort. For coding agents, OpenAI reports Terra solved 50.3 percent of Terminal-Bench 4.0 tasks at maximum effort, up from 21.5 percent for its predecessor, while nearly matching Anthropic’s Claude Sonnet 5.5 at medium effort. OpenAI also says Terra is more likely to stay within the scope it was given, to say plainly when it hits a limit, and to resist instructions designed to divert it. The company is honest about the ceiling, too: for the most complex coding tasks, it still points users to GPT-6.1 Sol and GPT-6 Astra.

Why computer use changes the job

Most software was built for human eyes and human hands. That is why so much office work still means opening one program, reading a screen, and typing what you found into another program. An agent that can see the screen skips the need for every tool to be rebuilt or connected in advance. It works with the software a business already owns, in the order a person would use it. That is a quieter kind of progress than a flashy demo, and probably a more durable one.

There is a practical takeaway for teams watching this space, including readers who enjoyed our look at Zwag AI putting every language first in Dubai. Start with work that is repetitive, easy to check and easy to undo: forms, invoice entry, first drafts of proposals, tidying drawings into models. Keep a person reviewing the output at first. The gains OpenAI is claiming, speed at half the price with fewer errors, are exactly the gains that show up in those jobs first.

Terra will not be the last word on agents, and its maker’s scores will get their outside tests soon enough. But the direction is now hard to miss. The assistant that only talks is giving way to the assistant that does. On the evidence of this launch, it does it faster, for less money, and with a steadier hand than the model it replaces. That is a good week for anyone with a screen full of small jobs waiting.

Also published on: