GPT-6.1 Terra by OpenAI is a medium-sized model with one unique ability: operating a computer like a human would. It looks at the screen, uses various applications, fills out forms, and composes a response, all at a very affordable cost compared to other models from the same company. The model was launched on October 26 and it is designed to handle the mundane tasks of office work, instead of groundbreaking studies. Those interested in the development of AI assistants will see if operating a computer can ever become an affordable task for AI agents. Here are the capabilities of the new model.
Also read: How do smart rings work? Sensors, accuracy and how they compare to smartwatches
The Terra model is the intermediate one of OpenAI’s current range. According to OpenAI, it features sophisticated visual reasoning and fine-tuned computer manipulation with enhanced writing style and is significantly advanced over GPT-5.6 Terra. The purpose of the model is performing unexciting tasks such as writing proposals, analyzing technical drawings, completing forms and editing video materials.
Computer manipulation involves the model looking at the user’s screen and clicking and typing as the user does. According to the OpenAI’s examples, these include turning architectural drawings into 3D models, or going through an animated movie frame-by-frame and preparing a time-coded list for restoration work. According to developer documentation, other examples are adding product pictures to catalog items and reacting to applications screens.
Besides, Terra is included into ChatGPT Work and Codex for Pro, Enterprise and Edu subscriptions, and is available for developers as gpt-6.1-terra. In addition, Terra supports skills, MCP and tool search functionality, which is important for agents having to act across multiple apps.
Also read: India’s affordable broadband comes at a cost: Excitel admits routers are falling behind
The first assertion is efficiency. As measured on the Agents’ Last Exam, a measure of professional abilities, Terra at medium-level reasoning ability achieves 15 percentage points more than GPT-5.6 Terra, but at about one-third of the price. OSWorld 2.1, a test of multi-step activities within desktop software, outperforms its predecessor’s best performance at a much lower price.
There is even greater efficiency in coding. As evaluated on Terminal-Bench 4.0, Terra at maximum level reasoning is able to solve 50.3 percent of tasks compared to 21.5 percent of GPT-5.6 Terra. OpenAI additionally claims that low-level reasoning error rate of 53.6 percent was reduced to 16.7 percent in its internal assessment, which is important when the model is representing you.
There is pricing to match. It is $1 per million input tokens and $5 per million output tokens, compared to $2 and $12 of the former version.
Each one of these figures was derived from OpenAI’s internal tests, and we will have to wait for some time until there are some independent tests. The claims should be regarded as just a starting point rather than the conclusion especially when it comes to cost comparisons depending on the degree of required reasoning.
However, a far more serious warning is related to behaviour of the software. Based on the reports regarding the system card of OpenAI, it turned out that the restriction imposed by the developers were bypassed by the system to achieve the goal in roughly 30% of all simulated tests. This might be surprising for the case of chatbot but represents a certain problem in terms of software which will be performing tasks on real accounts, filling various forms and making payments.
In any case, the trend seems to be quite clear. It looks like Terra confirms that using computer is becoming something cheaper than just a demonstration and more expensive than simple toy for playing, and the competition between AI laboratories shifts from providing the best response to accomplishing the task.
Also read: Air purifier myths: Don’t wait for bad AQI to switch it on in India