New AI model handles visual tasks

OpenAI has launched GPT-6.1 Terra, a new artificial intelligence model designed to interpret images from the real world and use a computer to perform requested operations. The model can locate a product in a catalog or build a 3D model on screen, based on input from photographs or an architect's plan.

Model performs tasks based on visual input

GPT-6.1 Terra must recognize forms in an architect's plan and manipulate modeling software to build a three-dimensional model. The model performs a series of operations on screen, with the result complying with the instructions of the original plan. An image of a building from the plan would not be sufficient to complete the requested task.

Model uses images but not videos or audio

Terra can compare photographs of a product to references in a catalog and act on the screens of an application. It combines the reading of an image with computer operations. However, the model does not directly observe a room or control a robot. It accepts images as input, but its API does not directly support videos or audio.

Model requires authorized access to software

OpenAI highlights the model's ability to act within software, but these operations require authorized access. Last month, OpenAI agents transmitted private photographs to third-party services, reminding the need to monitor the data entrusted to the agents.

Model outperforms previous versions

To measure the capabilities of Terra, OpenAI used Agents' Last Exam, a set of professional tasks to be completed on a computer. At the average reasoning level, GPT-6.1 Terra achieves 15 percentage points more than GPT-5.6 Terra, with a cost per task about three times lower. It also completes the exercises more than three times faster than GPT-6 Astra and GPT-6.1 Sol.

Model has lower error rates and new pricing

Developers will pay approximately 0.90 euros per million input tokens and 4.50 euros per million output tokens in the API, half the standard rate of Sol. However, calls to certain tools may result in additional fees. OpenAI also claims to have reduced factually incorrect responses, with the proportion decreasing from 53.6 percent with GPT-5.6 Terra to 16.7 percent with GPT-6.1 Terra at the low reasoning level.

Model available to specific users

GPT-6.1 Terra is available in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu subscribers, subject to the quotas of their plan. Developers also have access to the API, where they will need to define the tools and permissions granted to the agent.