NederlandsNew mid-sized OpenAI model operates the computer by itself, at half the price
OpenAI has released GPT-6.1 Terra, the second-smallest model in its range. It operates computer programs better than its predecessor, makes up answers far less often and costs less than a third per task. For now it is only available to paying users.
OpenAI released GPT-6.1 Terra on 26 October. OpenAI makes its models in four sizes: Luna is the smallest and cheapest, followed by Terra and Sol, and Astra is the largest and most powerful. Terra is therefore the second-smallest model. In the previous round of new models Terra was skipped; the latest version was still GPT-5.6 Terra from this summer.
The new Terra is mainly made for work on the computer: looking at what is on the screen, clicking, typing and switching between programs, the way a person at a laptop does. OpenAI says the model also writes more clearly and naturally and makes fewer mistakes.
What the model does better
The biggest difference from the previous Terra is in independent computer work. This is measured with OSWorld, a test in which an AI carries out tasks of several steps in ordinary programs, such as a browser, a word processor or a spreadsheet. A model can think for a shorter or longer time on each task: thinking longer often gives a better answer, but takes more computing work and therefore more money. According to OpenAI, the new Terra on a medium thinking setting already scores higher than the old Terra on its best setting. When the new model thinks longest, it solves 74 per cent of the tasks according to OpenAI, compared with 57 per cent for the previous version.
The model also makes far fewer mistakes on questions designed to trip it up. In OpenAI's own test, the error rate on the fastest setting fell from almost 54 to about 17 per cent. When the model thinks longest, it gets it wrong in just over 4 per cent of cases. That test is not public, so others cannot repeat it themselves.
Independent measurement: far fewer made-up answers
The benchmarking firm Artificial Analysis tests AI models itself, always with the same set of tests. The new Terra scores 48 points there, 6 more than its predecessor. That is less than OpenAI's strongest models: GPT-6.1 Sol is at 52 and GPT-6 Astra at 53. But per task in the test set, Terra costs 40 US cents, compared with 1.40 dollars for the old Terra. That is less than a third.
The most striking result is about honesty. Artificial Analysis also asks models questions they do not know the answer to, and checks whether they then admit they do not know or make something up anyway. The old Terra guessed in 88 per cent of such cases. The new Terra still does so in 38 per cent. That is a big step forward. Because the model still makes up an answer in more than a third of those cases, it remains wise to check important answers.
Last week the same ranking showed that the Chinese Kimi K3.1 Pro took fourth place, with 55 points. That model also improved most in independent work on the computer. Kimi scores higher, but a task from the test set costs almost 2 dollars there, five times as much as with Terra. The two show from different sides what AI makers are now focusing on: models that do not just talk, but also finish tasks on the computer themselves. Moonshot AI mainly went for the highest score, OpenAI with Terra mainly for a low price.
What difference does it make?
Anyone with a paid ChatGPT subscription (Plus, Pro, or a version for businesses or education) can now use the new Terra in ChatGPT Work. That is the part of ChatGPT that carries out tasks on its own, such as making a spreadsheet or presentation from your own files. On the computer it can also operate programs itself, and it asks for permission at important steps. Programmers can also use the model in Codex, OpenAI's programming tool. OpenAI does not mention free ChatGPT users.
For companies that build the model into their own apps, the price roughly halves. They pay 1 dollar per million tokens (pieces of text a few letters long) that they send to the model, and 5 dollars per million tokens the model writes back. For the old Terra that was 2 and 12 dollars. This makes it cheaper to let an AI do long tasks on its own, such as filling in forms or moving data between programs. Those are exactly the tasks in which a model takes many steps and therefore processes a lot of text.
Caveats
The figures on computer work and fewer mistakes come from OpenAI itself. Artificial Analysis confirms that the model has become cheaper and more honest; researchers can now check the other claims for themselves. Terra is also not the strongest model: for the hardest tasks, the larger models from OpenAI and its competitors remain better. What is new is that for a lot of ordinary work, the most expensive model is no longer needed.
What does this mean for the road to the singularity?
An AI that can operate a computer by itself can take over much more work than an AI that only answers questions. Until now this was mainly the domain of the largest and most expensive models. With the new Terra, OpenAI shows that a smaller model can do much of that work too, for a fraction of the price.
We are seeing this pattern more and more often: what only the top models could do a few months ago is now within reach of cheaper models. The lower the price per task, the more people and companies will let an AI assistant do real work on its own. The next step is for such models to become available to free users as well, and for them to indicate even more reliably when they do not know something.
What is the technological singularity? Read the explanation →
