What is GPT-6.1 Terra, and when should you use a cheaper computer-use model instead of a flagship?

GPT-6.1 Terra is OpenAI’s mid-priced model for writing, visual reasoning and routine computer use, introduced October 26, 2026 through the API and eligible paid ChatGPT Work / Codex access — not ordinary Chat, and API billing is separate. At $1/$5 per million tokens it is built for high-volume screen work. Here is the cost-versus-capability trade-off, how to test a routine task safely, and why it must never get your passwords.

Short answer: GPT-6.1 Terra is the model OpenAI introduced on October 26, 2026 for everyday knowledge work — clear, natural writing, visual reasoning, and routine computer use, where the model looks at your screen and clicks, types and switches apps for you. It is available through the API and through eligible paid ChatGPT Work and Codex access — not in ordinary ChatGPT chat — and API use is billed separately from any subscription. OpenAI prices it at $1 per million input tokens and $5 per million output tokens, half the list price of its GPT-6.1 Sol tier.Why it matters: computer use burns tokens fast.

An agent works in a loop — screenshot, decide, click, screenshot again — so a forty-step task pays for forty rounds of thinking. A model that is a little less brilliant but half the price can finish routine work for a fraction of the cost, and OpenAI says Terra matches its flagship Astra on its professional computer-use benchmark at roughly a tenth of the per-token cost (those are OpenAI’s own reported figures). Save the flagship for work where one wrong judgement is expensive: hard debugging or a final legal review.How to test a routine task safely, step by step:1.

Pick a boring, repeatable job first: entering invoice figures into accounting software, filling a standard web form, or drafting a routine proposal from a template.2. Use copies, not originals: point the agent at a test folder and a low-privilege test account, never your live books or your main login.3. Watch the first runs end to end: note where it hesitates or guesses; those spots need clearer written instructions, not a smarter model.4.

Set a spending cap in the API dashboard before you scale up, and judge cost per finished task, not per token.5. Run the same task once on the flagship: keep Terra only if the results are genuinely equal for that job.Pro tip: the common mistake is giving a computer-use model your passwords ‘to save time’. A model that can see your screen can see a password on it — approve logins yourself, every time.Ad’s take: a good cheap wrench handles the everyday bolts; the precision tool comes out for the cuts that count.

Terra is the everyday wrench.