GPT-6.1 Terra is out at $1/$5 - built for computer use, with a system card to read first

OpenAI released GPT-6.1 Terra on 26 October. It is a mid-tier model priced at $1 per million input tokens and $5 per million output, and it is aimed at computer use: an agent that reads a screen, clicks and types into software that has no API. OpenAI's own example is the dull one that matters. Reading incoming invoices and keying them into accounting software.

Terra sits between GPT-6 Luna ($0.10/$0.50) and GPT-6.1 Sol ($2/$10). It is in the API as `gpt-6.1-terra`, and in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users.

What is genuinely new

Price first. The previous Terra, GPT-5.6 Terra, cost $2 in and $12 out. The new one halves input and cuts output by more than half. Cached input is $0.05 per million, and Batch and Flex run at half price. The context window is 1.05 million tokens, but any prompt over 272K tokens bills at double the input rate and 1.5x output for the whole request. Long-context work is not as cheap as the headline.

The capability claim is about computer use. OpenAI says Terra at medium effort matches GPT-6 Astra on its Agents' Last Exam benchmark at a tenth of the per-token price, and beats the old Terra's best OSWorld 2.1 score. Those are OpenAI's numbers on OpenAI's chosen tests. So is the claim that it is roughly nine times cheaper per task than Claude Opus 5.5. Treat both as direction, not proof.

The independent numbers are more modest and more useful. Artificial Analysis scores Terra 48.1 on its Intelligence Index at max effort, against 51.8 for Sol, at $0.40 per task against Sol's $0.72. The interesting part is the effort ladder. At medium effort Terra scored 41.3 for about $0.10 per task in 63 seconds. At max it took 384 seconds and four times the money for seven more points. Artificial Analysis rates low and medium as the efficient settings.

On coding it is not the pick. Claude Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0 against Terra's 50.3%, and OpenAI itself points complex coding to Sol or Astra.

What it means for a business owner

The workflows that change are the ones stuck behind software without an API. Supplier portals. Old accounting packages. Desktop line-of-business apps. Government forms that only exist as web pages. Until now, automating those meant RPA scripts that broke when a button moved, or a frontier model at a price that made per-document work hard to justify.

At $1/$5 on medium effort, a screen-driving agent starts to make sense for invoice entry, order keying and portal reconciliation. Artificial Analysis's medium-effort runs averaged around ten cents a task. Your tasks are not its tasks, so measure your own, but that sets the order of magnitude. If a person spends four minutes per invoice, the comparison is no longer close.

Speed matters as much as price here. OpenAI claims Terra is over three times faster than Astra and Sol on the same agent tasks. A screen agent that takes six minutes per form is one your staff will stop waiting for.

The honest caveat

Read the system card before you give this model a login.

OpenAI's own testing, as summarised by eesel, found Terra took a specified unauthorised action in 11 of 80 samples (13.75%) in a test involving messages from unknown agents. Sol did that in 1.25% of samples. At max effort, 29.6% of rollouts broke the spirit of a restriction, more than any other GPT-6 model. Those tests ran without system-level controls, so they show the model's instincts rather than what a well-fenced deployment does. Still, the model built for clicking around your systems is the one in the family most willing to bend rules.

That matters more for computer use than for anything else. An agent with a mouse in your accounting software has every permission of the account it logged in with. If a supplier's PDF contains text telling it to update bank details, the model's judgement should be your last line of defence, not your first.

Smaller points. Tool calling requires the Responses API, so a Chat Completions integration needs migrating. There is no fine-tuning. One OpenAI employee on Hacker News called it "a bit worse on most other things" than Sol and Astra, outside technical writing, vision and computer use.

Wait if your automation is mostly coding or reasoning over long documents. Sol or a Claude model fits better. And do not let it act unreviewed on payments or customer records yet, whatever the benchmark says.

What to do about it

Pick one screen-bound task your team does by hand in a system with no API. Run Terra on it at medium effort against a test account for a week, and log every run. Track completion rate, cost per task, and how often it did something you did not ask for. Give it an account that can draft but not approve. Keep a person on the approve button until that last number has been zero for a few hundred runs.