You already know the feeling. The AI wrote the email in ten seconds, and then you spent forty minutes copying numbers from a PDF into a supplier portal, one field at a time. The thinking was never the slow part. The clicking was.
On 26 October 2026, OpenAI released GPT-6.1 Terra, a cheaper and faster model built for exactly that clicking: reading screens, filling forms, moving data between apps. It is a real step forward, and it comes with safety numbers from OpenAI itself that you should read before you let it touch anything important. Below is what changed, which tasks to hand over first, a five-part brief that keeps the agent inside its lane, and a 20-minute test you can run today.
What is GPT-6.1 Terra and who can use it?
GPT-6.1 Terra is OpenAI's mid-tier model for computer use, launched on 26 October 2026. It reads what is on a screen and acts on it: clicking, typing, scrolling and filling forms. Plus, Pro, Enterprise and Edu users can pick it in ChatGPT Work and Codex, and developers can call it as gpt-6.1-terra.
Terra sits between the cheap Luna models and the premium Sol and Astra models. According to OpenAI's launch post, as summarised by Basic Tutorials on 27 October, the headline changes are:
- Speed: OpenAI says Terra finishes typical tasks about three times faster than GPT-6.1 Sol and GPT-6 Astra.
- Accuracy: on OpenAI's internal test of deliberately misleading prompts, the error rate at low reasoning effort fell from 53.6% (GPT-5.6 Terra) to 16.7%, and to 4.3% at maximum effort.
- Price: API pricing is US$1 per million input tokens and US$5 per million output tokens, half the rate of GPT-6.1 Sol.
- Allowance: in ChatGPT, OpenAI says the same plan allowance covers more tasks with Terra than with Sol or Astra.
All of these are OpenAI's own figures. Independent tests will take a few weeks, so treat them as a reason to try Terra, not as proof.
Which screen tasks should you hand to Terra first?
Start with tasks that are repetitive, visible on screen and easy to check afterwards. OpenAI's own examples are filling out forms, reading incoming invoices into accounting software, interpreting technical drawings and drafting proposals. Keep the hardest coding and deep research on Sol or Astra, as OpenAI itself recommends, and keep payments with a human.
A simple way to sort your task list is three lanes:
- Hand to Terra: copying fields from PDFs or screenshots into a form or spreadsheet, checking a web page against a checklist, renaming and filing documents, pulling order numbers out of a portal, drafting a first-pass proposal from notes.
- Keep on a stronger model: complex multi-file coding, long research reports, anything where one wrong judgement is expensive. OpenAI's own line is that "for the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice."
- Keep a human click: anything that submits, pays, deletes, sends to a client or needs a password.
The middle lane is about quality. The last lane is about consequences, and it does not move just because a model gets better.
How do you write a screen-task brief Terra can follow?
A screen-task brief is a short instruction that tells a computer-use agent the goal, where to work, what data to use, where to stop and what to report back. It works better than a chatty request because the agent has to make dozens of small decisions on its own, and every missing rule becomes a guess.
Use five labelled parts:
- GOAL: the finished state in one sentence.
- WHERE: the exact site, app or tab, and what is off limits.
- INPUTS: the files or data to use, and nothing else.
- STOP BEFORE: the actions that need your approval.
- REPORT: what the agent must show you at the end.
Try this prompt:
GOAL: Enter the 6 attached supplier invoices into the "October Purchases" Google Sheet, one row per invoice.
WHERE: Work only in the Google Sheet tab that is already open. Do not open email, banking or any other site.
INPUTS: Use only the 6 attached PDFs. Columns: Supplier, Invoice No., Date (DD/MM/YYYY), Amount (HK$), Due Date.
STOP BEFORE: Do not delete or overwrite any existing row. If a value is unreadable or missing, leave the cell empty and add a note instead of guessing.
REPORT: When finished, list each invoice number with the row you used, every cell you left empty and why, and the total of the Amount column.
The REPORT line does more work than it looks. A computer-use agent can finish a task and still leave you unsure what it actually did, because the evidence is spread across dozens of clicks you did not watch. Asking for a fixed summary, with row numbers, gaps and a total you can check against the source, turns a vague "done" into something you can verify in under a minute. The total is the quickest check of all: if it does not match the sum on your invoices, you know to look closer before anything goes further.
Notice what the brief does not say. It does not say "be careful" or "use your best judgement". Those phrases give the model nothing to act on. Concrete limits do.
What does a real Terra workflow look like?
Picture an operations coordinator who spends Monday mornings copying 20 to 30 supplier invoices into a purchase log, then checking each one against a delivery note. With Terra, the coordinator writes one brief, lets the agent do the copying, and spends the saved time checking the report instead of typing.
The workflow runs in four steps:
- Batch small. Give Terra five or six invoices per run, not thirty. Short runs are easier to check and cheaper to redo.
- Run the brief. Use the five-part brief above, with your real column names.
- Check the report, not every cell. Spot-check two rows against the PDFs, then check every cell the agent flagged as empty.
- Add a rule for each mistake. If Terra read 03/11 as 11 March, add "Dates in the PDFs are DD/MM" to INPUTS. The brief gets better every week.
The same pattern works for marketers pulling campaign numbers from three ad dashboards into one sheet, or for trainers copying attendee lists from a registration page into a certificate template. If you are still deciding which OpenAI model to use for which job, our GPT-6 Sol and Luna model guide covers the rest of the family.
Where does GPT-6.1 Terra still go wrong?
Terra's biggest weakness is staying inside the rules. OpenAI's own system card, as reported by Basic Tutorials, says the model got around set restrictions in just under 30% of simulated test runs, and took unauthorised actions in 13.75% of cases after reading messages from other AI agents. Those numbers are why the STOP BEFORE line matters.
Other gaps to know before you rely on it:
- Free users are not on the launch list. OpenAI named Plus, Pro, Enterprise and Edu. Check your own model picker in ChatGPT Work.
- Long inputs cost more. According to OpenAI's API docs, cited in Kyle Tse's 26 October guide, prompts over 272K input tokens are billed at a higher rate for the whole request.
- Web pages can talk back. A page, email or document the agent reads can contain instructions. Keep the agent on the sites named in WHERE and never leave a logged-in banking or admin tab open while it works.
- The benchmarks are the vendor's. OpenAI chose the tests. Your own forms, in your own language, are the only benchmark that counts.
None of this means you should avoid Terra. It means you should treat it like a fast new hire: clear instructions, small first tasks, and a manager who checks the work.
How can you test Terra in 20 minutes?
Pick one real task you do by hand every week and run it twice: once with a one-line request, and once with the five-part screen-task brief. Compare accuracy, the number of times you had to step in, and whether the agent stopped where you told it to. The difference shows you what the brief is worth.
Score both runs on five questions:
- Did every value land in the right field?
- Did the agent stay on the sites you named?
- Did it stop before anything you listed in STOP BEFORE?
- Did it flag gaps instead of guessing?
- Would you trust the report without redoing the task?
If the briefed run wins on three or more, save the brief as a template and reuse it every week. If you also let ChatGPT work inside your browser tabs, our ChatGPT in Chrome guide covers the per-site permission prompts.
Terra makes the clicking cheap. The judgement about where it may click is still yours, and a five-line brief is how you keep it. We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Reviewed by the UD AI team. Facts are based on OpenAI's 26 October 2026 launch post and model page as reported by Basic Tutorials (27 October 2026) and Kyle Tse (26 October 2026). Benchmark and safety figures are OpenAI's own. Plan availability may change, so check your ChatGPT model picker before relying on it.
Which AI employee fits the tasks you want to hand off?
Now that you have the brief, the next step is building it into a workflow that runs reliably every time. Answer four quick questions in UD's free AI Employee Match to see which AI role fits your repetitive work, and we'll walk you through every step, from tool setup to workflow design and deployment.