Skip to content
Bamboo Flags
Custom automation & AI for growing businessesDiscover · Design · Deliver · Drive

GPT-6.1 Terra: OpenAI's cheaper model for computer use tasks

On 26 October 2026 OpenAI released GPT-6.1 Terra at $1 and $5 per million tokens. This article lists what it costs, what is not known and the checks to run.

Bamboo Flags3 min read

On 26 October 2026, OpenAI released GPT-6.1 Terra, a mid-priced model that it presents with the line "Bringing fast and capable computer use to everyone". Computer use means that the model operates a browser or a desktop: it reads screenshots and returns clicks and key presses.

For a small business, this matters for screen work in a web tool that has no API. Terra makes this type of automation cheaper. The safety rules for computer use do not change, and some facts are not known yet.

What GPT-6.1 Terra costs

The model page gives these prices per million tokens: $1 for input, $0.05 for cached input and $5 for output. A request with more than 272,000 input tokens costs two times more on input and 1.5 times more on output.

The Register lists the OpenAI range, input and output per million tokens:

  • GPT-6 Luna: $0.10 and $0.50
  • GPT-6.1 Terra: $1 and $5
  • GPT-6.1 Sol: $2 and $10
  • GPT-6 Astra: $10 and $50

According to OpenAI, Terra costs 50% less than GPT-5.6 Terra and GPT-6.1 Sol. For example, a run with 200,000 input tokens and 10,000 output tokens costs $0.25 at standard prices. The real count depends on the number of screenshots and steps.

What the model can do

The model page lists computer use as a supported tool, together with web search, file search, a code interpreter and MCP. Tools work only through the Responses API. The context window is 1,050,000 tokens and the knowledge cutoff is 1 June 2026, later than the 30 April 2026 cutoff of Astra and Sol.

The model has six reasoning levels, from none to max, and the level changes cost and time. Simon Willison tested Terra at each level with one drawing task. At max, it cost 11.3 cents and took 4 minutes and 11 seconds. Willison also notes that computer use suits a cheaper model, because it is often repetitive work.

Ted Sanders, an OpenAI engineer, wrote on Hacker News that Terra is similar to Astra and Sol on vision and computer use, but much faster, and much smarter than Luna. He also wrote that it is "a bit worse on most other things".

What is not known

  • The reports we read give no benchmark numbers for computer use. The comparisons above are informal comments from an OpenAI engineer.
  • OpenAI's computer use guide does not mention Terra on 27 October. Its examples use Sol and Astra.
  • According to The Register, Terra refuses harmful requests, and unnecessary ones, less often than Sol. No source says what this means for an agent on a website.

The safety rules stay the same

OpenAI's guide gives five controls. Use "an isolated browser or VM and an allow list of sites and actions." Treat text on the screen as untrusted: it "cannot grant permission or override the user's instructions." Keep a person in control of purchases, data transmission and changes that are hard to reverse. Remember that "typing sensitive information into a form counts as transmission." Set limits on each run and check the actual result, not only the model's final answer.

Checks before a pilot

  1. Pick one repetitive screen task in a tool without an API. If the tool has an API, use the API. Some websites also block AI agents.
  2. Run the task in a separate browser profile or virtual machine, with no saved passwords.
  3. Write an allow list of the sites the agent can open.
  4. Make the agent stop and ask before it pays, sends data or deletes something.
  5. Start at a low reasoning level and record the tokens and the time of each run.
  6. For the first two weeks, compare each result with the expected result.
  7. Decide a monthly budget for the pilot before the first run.

Found something worth fixing?

Bring it to a free 45-minute call.

Book a free 45-minute call

More from the blog