AI Can Fill In Your Forms Now. Decide Which Boxes It Never Touches.
OpenAI's new GPT-6.1 Terra is built to read the screen and fill out forms. Here is how to hand it the paperwork without handing it the signature.
By John Martines · November 1, 2026 · 6 min
OpenAI's GPT-6.1 Terra is aimed squarely at the forms and paperwork that eat an office manager's afternoon. Here is what it does, what its own error numbers mean, and the "never" list to write before anyone lets it type.
A new commercial client signs with a cleaning company on Monday. By Tuesday their purchasing team has sent a vendor setup form: business name, address, insurance details, service locations, three contact names, a W-9 to attach. The office manager has filled in forms like it fifty times. It still takes half an hour of copying from old files.
This week she could point an AI at the form and at the folder with last year's copy, and watch it fill in the boxes. That's the pitch behind the model OpenAI released on October 26, 2026. It's a real time saver. It also needs one page of rules before anyone lets it type.
GPT-6.1 Terra is OpenAI's new middle model: smarter than its cheapest, cheaper than its best. What's new is the job it's built for. OpenAI says it "excels at the communication and computer-based tasks that fill the workday," and names drafting proposals, reading technical drawings and filling out forms.
It does that through computer use. The AI looks at the screen the way you do, reads what's there, and clicks and types in the boxes. It isn't connecting to the form's software behind the scenes. It's doing the copying a person would do, only faster.
It runs in ChatGPT Work, the part of ChatGPT that works across your files and desktop apps, and in Codex, OpenAI's coding tool, not in the regular chat window. Coverage lists Plus, Pro, Enterprise and Edu plans, and access depends on your plan, region and how far the rollout has reached, so you may not see it yet. For developers, OpenAI cut the price to half of what the Terra it replaces cost.
How it got here:
- July 17, 2025: OpenAI launched ChatGPT agent, which could use its own browser to fill forms and was built to ask permission before actions of consequence.
- September 2026: OpenAI released its GPT-6 Sol and Luna models, leaving the middle model, Terra, on the older version.
- October 26, 2026: GPT-6.1 Terra arrived in ChatGPT Work, Codex and the developer tools, pitched for the whole workday at half the old price.
Most small offices don't have a paperwork problem. They have a copying problem. Vendor setup forms, insurance certificates, permit applications, supplier credit applications, patient intake systems, job bid portals. The information already exists somewhere in the building. Someone spends the afternoon moving it from one box to another.
A model aimed at exactly that work, inside a plan you may already pay for, is worth trying. But read OpenAI's own numbers carefully.
It still makes mistakes, by its maker's count. On OpenAI's internal set of questions built to trip a model up, Terra got facts wrong 16.7% of the time at its quickest setting and 4.3% at its most careful one. That's a big improvement on the version before. It's also a test designed to be hard, not a forecast of how often your form will be wrong. The honest reading: it will get some boxes wrong, and you won't know which ones unless someone checks.
A wrong box on a form looks exactly like a right one. A draft email with a mistake reads oddly. A form with one wrong digit in a policy number reads perfectly. The more polished the rest of the form, the easier the bad box is to miss.
Some boxes are promises. A delivery date, a price, a bank account for payment, a signature, a checkbox that says "I certify." Those aren't typing. They commit the business, and they belong to a person.
1. Write the "never" list first. One page, taped next to the computer. Things the AI never fills in on its own: bank account and routing numbers, prices and rates, dates you're promising, anything that signs, certifies or agrees to terms on the business's behalf. Paste the list into your request every time, so the AI sees it too.
2. Pick one boring, repeat form. Choose a form you fill in every month and know by heart, like a supplier reorder sheet or a client's vendor update. You'll spot its mistakes because you know the right answers.
3. Point it at the source, not at its memory. Give it the folder with last year's copy, your insurance certificate and your company details sheet. An AI working from your documents copies. An AI working from nothing guesses, and a guess looks just as confident.
4. Let it fill, then stop. Tell it plainly: fill in the form, don't submit it. A person then reads every field against the source paperwork, fixes what's wrong, and presses submit. The AI did the typing. A person signs for what the form says.
5. Keep a short error log for a month. Every time you fix a box, write down which box and what it got wrong. After a month you'll know which forms it handles well and which boxes it trips on. If you want to build the checking habit first, our free [Check-the-Work Trainer](/tools/trainer) gives you AI drafts with hidden mistakes to find.
- Is there anything on this form from the "never" list? If so, a person fills those boxes.
- Did you give it the source documents, or is it working from memory?
- Is it set to stop before submitting, and have you seen it stop?
- Who reads every field against the paperwork, and do they have the time to do it properly?
- If a wrong form goes out, who notices, and how fast can you correct it with the customer or agency?
AI is moving from writing about your paperwork to doing it. The models will get faster and cheaper, and more of them will read the screen and type into whatever form is in front of them. That's genuinely good news for an office manager who spends Tuesday afternoons copying insurance details. The businesses that decide which boxes the AI may fill, and who checks them, will get those afternoons back without the wrong numbers.
---
Sources (checked October 29, 2026). GPT-6.1 Terra launch, OpenAI's description of its workday tasks, availability in ChatGPT Work, Codex and the API, the 50% lower price, and the 16.7% and 4.3% error rates on OpenAI's internal benchmark: "OpenAI upgrades ChatGPT and Codex with new GPT-6.1 Terra model" (9to5Mac, Oct 26, 2026), quoting OpenAI's announcement. Computer use, plans and rollout: SiliconANGLE and The Next Web coverage (Oct 26, 2026) and OpenAI's ChatGPT release notes (Oct 2026). Note on what the error benchmark measures: Artificial Analysis, "GPT-6.1 Terra pushes the cost efficiency frontier" (Oct 2026). ChatGPT agent and its permission prompts: OpenAI, "Introducing ChatGPT agent" (July 17, 2025).