Analysis
OpenAI's documentation describes GPT-6.1 Terra as a fast, efficient model for visual reasoning, computer use and natural writing. Its examples include matching product photographs with catalogue entries, interpreting technical drawings and acting on application screens. The model page supports those intended uses; it does not establish when Terra launched.
OpenAI's model-selection guidance also points to tasks with a defined outcome and supplied material. That is a useful distinction for a screen-operating assistant: reading a page, choosing a control and saving the intended result are separate parts of the job. A convincing explanation of what happened is not evidence that the final record is correct.
Our reading: evaluate Terra on a complete workflow with a known starting state and a result that can be reopened. Count mistaken edits, retries and human checking alongside elapsed time. A model that responds quickly can still make a slow assistant if someone must repair its work. We have not independently measured Terra's speed, accuracy or cost advantage.
Source: GPT-6.1 Terra Model | OpenAI API
Why it matters
Computer-use models need to be judged on saved outcomes, including the effort required to check them.