On October 26, 2026, OpenAI released GPT-6.1 Terra, a mid-priced model built to read screens, work inside apps and write. For a business owner, the headline is not the benchmark chart — it is that handing routine computer work to AI just got cheaper again. Here is what changed, what it costs, and what still needs a person.
What did OpenAI release on October 26?
OpenAI released GPT-6.1 Terra, which its developer documentation describes as "a fast, efficient model for visual reasoning, precise computer use, and clear, natural writing" (as of October 29, 2026). It replaces GPT-5.6 Terra and is priced between OpenAI's top models (GPT-6.1 Sol and GPT-6 Astra) and its low-cost Luna models.
Where you can use it, according to OpenAI's launch notes of October 26, 2026:
- ChatGPT Work — the part of ChatGPT that carries out longer, multi-step tasks — on Plus, Pro, Business, Enterprise and Edu plans, rolling out gradually.
- Codex, OpenAI's coding tool, on the same plans.
- The API (the paid connection developers use to build it into other software), as
gpt-6.1-terra.
It is not available in the regular ChatGPT chat window. Business members get it inside the Work and Codex seats they already have, subject to what their workspace admin allows.
How much cheaper is it?
GPT-6.1 Terra's list API price is $1 per million input tokens and $5 per million output tokens, against $2 and $12 for the GPT-5.6 Terra model it replaces, according to OpenAI's pricing page as of October 29, 2026. A token is a small chunk of text; OpenAI's rule of thumb is that 100 tokens is about 75 English words.
| Model | Input, $ per 1M tokens | Output, $ per 1M tokens | How OpenAI describes it |
|---|---|---|---|
| GPT-6 Astra | $10.00 | $50.00 | "Our most capable model for the most demanding work" |
| GPT-6.1 Sol | $2.00 | $10.00 | "Near-Astra performance for complex work at a lower cost" |
| GPT-5.6 Terra (replaced) | $2.00 | $12.00 | Previous mid-tier model |
| GPT-6.1 Terra | $1.00 | $5.00 | "Fast, efficient" — screens, apps, writing |
| GPT-6 Luna | $0.10 | $0.50 | "Our most efficient model for focused, high-volume tasks" |
Standard tier, short-context prices (up to 272,000 input tokens), OpenAI pricing page, October 29, 2026.
OpenAI also claims Terra performs on par with GPT-6 Astra at one-tenth of its price per token and finishes tasks more than three times faster than Astra and Sol. Those are OpenAI's own tests, not independent results.
What does a routine admin task cost now?
A short email task costs a fraction of a cent at API list prices. Our estimate, using OpenAI's 75-words-per-100-tokens rule: reading a 1,500-word email thread (about 2,000 tokens) and drafting a 150-word reply (about 200 tokens) costs roughly $0.003 on GPT-6.1 Terra, versus roughly $0.006 on GPT-5.6 Terra.
Real tasks cost more than that arithmetic suggests, for three reasons:
- Screens and files add tokens. OpenAI notes that images, files, tools and message structure all count toward the bill — and a model working inside apps is looking at images of the screen.
- Thinking adds tokens. Terra lets you choose how much it reasons before answering, from none up to max; more reasoning means more output tokens.
- Retries add tokens. A task that fails and reruns is paid for twice.
If you use ChatGPT rather than the API, you pay for your plan, not per token. OpenAI says Terra's usage allowance in ChatGPT Work goes further than GPT-5.6 Terra's.
What is "computer use," and why should an owner care?
Computer use means the model looks at a screen and clicks, types and moves between apps the way a person would. OpenAI's own guide gives examples such as acting "on application screens" and matching "product photos to catalog entries."
That matters because many small businesses run on software that does not connect to anything else: supplier portals, insurer portals, older accounting tools, government websites. Until now, getting work done in those systems meant a person copying and pasting. Tasks worth testing:
- Copying invoice details from emailed PDFs into a billing portal.
- Checking order status on several supplier sites and updating one spreadsheet.
- Updating customer records in a CRM (customer relationship management system) after calls.
- Filling in the same details on several insurer or licensing portals.
What still needs a person?
Judgment, approvals and exceptions still need a person, and OpenAI's own figures show why. OpenAI says Terra's error rate on its internal set of prompts designed to trigger mistakes fell from 53.6% for GPT-5.6 Terra to 16.7%, at low reasoning effort (October 26, 2026 announcement) — a real improvement, but still about one wrong answer in six on hard cases.
Keep a person on:
- Anything that moves money or makes a promise — payments, refunds, quotes, contract terms.
- Exceptions — disputes, upset customers, requests that do not fit the usual pattern.
- Facts that matter — totals, dates, names, account numbers.
- The rules themselves — deciding what "done" looks like for each task.
OpenAI itself points users to Sol or Astra, not Terra, for in-depth research or complex coding.
What should you do this week?
Run one small, supervised test rather than a big rollout:
- Pick one repetitive computer task that takes someone at least two hours a week.
- Write down the steps and what a finished result looks like.
- If you are on Plus or Business, check whether Terra appears in ChatGPT Work — rollout is gradual, and Business admins control model access.
- Run it on ten past examples with a person checking every output.
- Count errors and minutes saved, then decide whether to continue.
- Check data settings first: OpenAI says Business workspace data is excluded from model training by default, while personal Free, Plus and Pro accounts share data unless the user turns it off.
Where does a managed AI worker fit?
A cheaper model lowers the cost of each task; it does not run the workflow for you. Someone still has to connect it to your systems, catch exceptions, keep a record of what it did, and fix it when a website or the model changes. You can do that yourself in ChatGPT Work, or have it run for you — we compare the two in managed AI vs DIY AI.
That second route is what we call an AI operator: one workflow, such as chasing unpaid invoices, run end to end, proven on your own data before you commit. Cheaper models like Terra make that kind of setup less expensive to run, whoever runs it.
If you want a structured way to pick your first workflow, the AI handoff playbook walks through it step by step. And if you would rather talk it through, a consultation is a good place to start.