Test GPT-6.1 Terra on one screen task before you scale it.

What OpenAI released

On Oct 26, 2026, OpenAI introduced GPT-6.1 Terra. OpenAI calls it "a fast, efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing."

OpenAI says it is "available to all Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex. Developers can also access it through the OpenAI API as gpt-6.1-terra."

What it is in plain words

GPT-6.1 Terra is built for everyday work done on a screen. It can look closely at images and documents. It can use software the way a person does, by clicking and typing. It also writes proposals and other text that read more naturally.

OpenAI's examples include drafting proposals, interpreting technical drawings, filling out forms, editing videos, and reading invoices before entering the details into accounting software.

The invoice example tells you who this is for. Most small businesses don't have a pile of hard research problems. They have a pile of copying information from one screen to another. That is the work this model is aimed at.

OpenAI's numbers, and what they are

All of these figures come from OpenAI's own launch materials. No independent testing is cited, so treat them as the vendor's claims.

If these numbers hold up, the model makes up fewer facts at low effort, handles multi-step technical work better, and costs less per computer task. Only you can find out whether they hold up on your invoices and your software.

Why it matters to a small business

There are three reasons.

1. 01

   Cost.

   OpenAI says API prices are 50% lower than standard prices for GPT-5.6 Terra and GPT-6.1 Sol. If you or a contractor build small automations on the API, it now costs less to try an idea.

2. 02

   Computer use.

   Many of the tools small businesses rely on have no clean way to connect to other software: the older accounting package, the supplier portal, the government form. A model that can work the screen directly can reach those places, as long as someone is watching.

3. 03

   Writing.

   Customer emails, quotes and follow-ups go out under your name. Clearer drafts mean less time spent editing before you hit send.

The catch is that a model that can click buttons can also click the wrong ones. When it's this fast and this cheap, it's tempting to let it run on its own. For now, don't.

For more on where agents fit in ChatGPT Work, see our related reading on ChatGPT Work agents for small business. For broader ideas, see our guide to practical GPT-6 uses.

Where to start, and where to keep a human

1. 01

   Invoice intake.

   Let the model read incoming invoices and draft the entry in your accounting software. A person approves each entry before it posts. No money moves until a human says yes.

2. 02

   Catalog cleanup.

   Have it match product photos to catalog entries and flag the mismatches. Your team reviews the flagged items instead of checking every listing.

3. 03

   Form prep.

   Use it to pre-fill repetitive forms such as supplier onboarding, warranty claims and permit applications. A person checks the form and submits it. Don't let the model submit anything that commits you legally.

4. 04

   Drawing reads.

   If you run a trade or manufacturing shop, ask it to read a technical drawing and list the dimensions or materials. Treat that as a first pass, and have the person who signs off on the job confirm it.

5. 05

   Customer drafts.

   The model drafts replies, quotes and follow-ups. A person reviews and sends them.

The same rule covers all five: the model prepares, and a person approves anything that touches money, customers or legal commitments. Give the model its own login with only the access it needs. Never give it your admin account.

Read with care

1. 01

   Availability.

   OpenAI names Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, plus developers on the API. The announcement says nothing about other plans. If you're on another plan, check your own account before you assume you have it.

2. 02

   Benchmarks.

   Every figure above is OpenAI's own. Vendor benchmarks are vendor claims, run on tests the vendor chose. The test that counts is your own workflow.

3. 03

   Cybersecurity rating.

   In its system card addendum, OpenAI rates GPT-6.1 Terra "Critical" in cybersecurity under its Preparedness Framework. It says the model keeps the protections it built for GPT-6 Astra. Put simply, OpenAI treats the model as highly capable at cybersecurity tasks and applies safeguards to match. For you, the practical step is the same as with any capable tool: give it limited access and keep a record of what it does.

This week

1. 01

   1. Pick one task.

   Choose a single screen task your team repeats every week, such as invoice entry or catalog matching. Write down how long it takes today.

2. 02

   2. Check your plan.

   Confirm whether GPT-6.1 Terra shows up in your ChatGPT Work account, or whether your developer can reach it through the API.

3. 03

   3. Run it side by side.

   Have the model work a small batch while the person who normally does the task checks every result. Note its mistakes as well as the time it saves.

4. 04

   4. Write the approval rule.

   In one sentence, state what the model may prepare and what a person must approve. Share it with your team before anyone uses the model on live work.

Our take

GPT-6.1 Terra is aimed at the screen work that eats up a small team's week, and OpenAI's pricing makes it cheap to find out whether it helps. Start with one task, keep a person on the approve button, and let your own results decide how far it goes, not the benchmarks.

Sources

Introducing GPT-6.1 Terra (OpenAI, Oct 26, 2026)

Addendum to GPT-6 Astra System Card: GPT-6.1 Terra (OpenAI, Oct 26, 2026)