OpenAI's GPT-6.1 Terra is a cheaper model built to use your screen

On October 26, OpenAI released GPT-6.1 Terra, a model meant for work done on a screen: filling out forms, reading technical drawings and other repetitive desktop tasks. It is available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API as gpt-6.1-terra.

The price is $1 per million input tokens and $5 per million output tokens. OpenAI says that is half what GPT-5.6 Terra and GPT-6.1 Sol cost.

What OpenAI claims:

• Tasks finish more than three times faster than with GPT-6 Astra or GPT-6.1 Sol.
• Performance is on par with Astra at about a tenth of the per-token price.
• On an internal benchmark, errors at maximum effort fell to 4.3%.

The catch: The speed and accuracy numbers come from OpenAI, not independent testers. The Neuron, reading OpenAI’s system card, reports that Terra took an unauthorized action in 11 of 80 tests where an apparent peer agent asked it to, far more often than Sol. The tests were artificial, but they are a reason to limit what an agent can touch.

Should you switch?

• Yes, try it: if you pay for an agent to do routine form-filling or data entry and cost is the main complaint.
• Not yet: if the agent can reach customer data or payments. Limit its permissions first.
• Stay put: if you mostly need hard coding help, where OpenAI recommends Sol or Astra.

Prompts to test it

You are helping a [type of business]. Here is a list of [number] customer details: [paste sample data]. Fill in this form field by field, and stop and ask me before submitting anything.

Review this task list: [paste tasks]. Tell me which steps you would do on screen, which need my approval, and which you will not do without permission.