OpenAI added another model to its October lineup. GPT-6.1 Terra arrived on October 26, 2026, targeting everyday professional tasks: visual reasoning, computer use and writing.

Where you can use it

  • In the API as gpt-6.1-terra.
  • In ChatGPT Work and Codex.
  • Not in the regular ChatGPT Chat tab.

Pricing

For a mid-tier model, Terra is cheap: a dollar buys a million tokens of input, cached input costs five cents per million, and output runs $5 per million. A ChatGPT subscription does not cover any of it; API calls are billed on their own.

Other October changes

  • October 14: GPT-5.5 was retired from ChatGPT, ChatGPT Work and Codex. API access is not affected.
  • October 30: GPT-6 Rosalind Research, a model for multistep life-sciences work with extra safeguards, began rolling out to approved research customers in Work, Codex and the API.

What “computer use” means

A computer-use model operates software the way a person does. It receives a screenshot of a screen, decides what to do next, and returns an action such as clicking a button, typing into a field or scrolling. The app running the model carries out that action, takes a new screenshot and sends it back. Repeat that loop dozens of times and the model can complete a task like filing an expense report in a web form that has no API.

That loop is why visual reasoning matters so much. The model has to read cluttered interfaces correctly, find the right button among many, and notice when a page didn’t load or an error message appeared.

What the pricing means in practice

AI models are billed per token, a chunk of text roughly three-quarters of an English word on average. Screenshots also count as input tokens. Because a computer-use agent sends a fresh screenshot and the conversation so far on every step, input tokens pile up much faster than output.

A quick example using Terra’s published rates: a task that takes 30 steps and uses 400,000 input tokens and 20,000 output tokens in total would cost about $0.40 for the input and $0.10 for the output, or roughly half a dollar per task. If much of the repeated context can be served from the cache at five cents per million tokens, the input part falls sharply. That is the kind of math teams running thousands of tasks a day will be doing.

Why it matters for developers

A cheaper model tuned for computer use is useful for agents that click through interfaces or fill in forms, where many calls add up quickly. If you build on the API, run Terra against your current model on your own tasks before switching.

When you test, measure three things rather than one: how often the agent completes the task without help, how many steps it takes, and the total cost per finished task. A model that is cheaper per token but needs twice as many steps, or fails more often, can end up costing more.

What we think

OpenAI’s October releases show the market splitting into tiers: flagship models for the chat experience, specialized models for research, and cheaper workhorses like Terra for automation. That is good news for businesses, which rarely need the most powerful model for routine work. Agents that operate real software also raise the stakes of mistakes, so we’d start with low-risk, easily reversed tasks and keep a person approving anything that sends money or messages.