GPT-6.1 Terra and supplier invoices: what to test first

Typing supplier invoices into accounting software is the kind of chore OpenAI now says its newest model can take over. GPT-6.1 Terra, released on October 26, 2026, is built to operate a computer through its screen, and OpenAI says its API prices are 50% lower than those of GPT-5.6 Terra. The company's own example of what that makes practical is reading incoming invoices and entering their details into accounting software. For a small business still keying invoices by hand into a system such as Priority (an Israeli business-management package) or SAP Business One, a test is worthwhile, provided a person approves every entry.

Who can use it

The model is rolling out to Plus, Pro, Enterprise and Edu subscribers in ChatGPT Work and Codex, according to OpenAI's developer forum announcement. Business plan customers are not on that list. Terra replaces GPT-5.6 Terra, and developers call it as gpt-6.1-terra.

ChatGPT subscribers should not expect a lower bill. The price cut covers the API only; what subscribers get instead is a usage allowance that stretches further. OpenAI says the allowance goes further with GPT-6.1 Terra than with GPT-5.6 Terra, and far further than with GPT-6.1 Sol or GPT-6 Astra, but it gives no figures, and no source says Plus got cheaper.

Developers pay $1 per million input tokens, $0.05 for cached input and $5 for output, Techstrong.ai reports. Tokens are the chunks of text that usage is billed by. GPT-5.6 Terra had already been cut in July to $2 for input and $12 for output, so the new rates are 50% lower on input and about 58% lower on output. The headline figure, in other words, is not uniform.

A model that works the screen

TechRepublic describes Terra as using a computer the way a person does: looking at the screen, clicking and typing. That lets it work inside accounting software through its normal interface, with no ready-made integration. Think of a temp at the keyboard, reading each invoice and keying in the fields.

OpenAI is candid that this is not its strongest all-round model. A quote in the forum post puts Terra on par with Sol and Astra for vision and computer use, "but much faster," and "a bit worse on most other things." For in-depth research or complex coding, OpenAI points users to Sol or Astra.

The invoice example also needs a careful read. On OpenAI's announcement page it describes developers automating high-volume work through the API, not a built-in ChatGPT Work feature for supplier invoices. Only an excerpt of that page was accessible.

The error rate

BleepingComputer reports that in OpenAI's internal test at low reasoning effort, the error rate fell from 53.6% with GPT-5.6 Terra to 16.7% with GPT-6.1 Terra. That is a steep improvement, yet it still works out to roughly one error in six, which matters when the output is financial data. The test was OpenAI's own and did not involve invoices; no benchmark was found for invoices, Hebrew-language documents or scans. TechRepublic adds that a lower error rate does not remove the need to verify sensitive financial actions.

What changes for a small business

Two routes seem likely. An owner on a Plus plan can try a handful of invoices at a time in ChatGPT Work. Dozens or hundreds of documents would probably call for a developer building a process on the API, which is where the price cut applies.

Consider a bookkeeping firm that receives supplier invoices from several clients by email. Instead of an employee typing each one, the model extracts supplier, invoice number, date and amount into a list, and the employee reviews, corrects and approves it. The work does not disappear; it moves from typing to checking.

Testing it without touching the books

1. Check the plan and open Work. OpenAI's help page says to open ChatGPT and choose "Work"; in the desktop app, the options are Work or Codex. The app may ask for screen recording and accessibility permissions. In an organization, an admin can turn on "Work Cloud" for members.
2. Start with a review queue. TechRepublic advises extracting invoice fields into a list that a person checks before anything is posted. A practical start is to upload a few supplier invoices, ask for a spreadsheet of the key fields and compare each row with the original.
3. Run a side-by-side test. The same invoices go through the current process and through the model, tracking completion rate, time, cost and reviewer effort. Before any move to the API, TechRepublic's warning applies: the token price is only part of the cost, alongside repeated model calls, integration work, human review and fixing mistakes.

Any agent that touches the accounts needs guardrails, according to TechRepublic: least-privilege access, human approval before consequential changes, logs of every action, and protection against prompt injection, meaning instructions hidden inside a document such as an invoice that hijack the agent.

No source addresses Hebrew support or availability in Israel. Nor does any say what happens to uploaded supplier data, so the terms of service deserve a look before real invoices go in.

The experiment makes little sense for a business that receives only a few invoices a month, or for one with nobody to check the output before it is posted. Business plan customers would also need to switch plans to try it in ChatGPT, which is only worth doing if there are other reasons to.

A short checklist

• The subscription is Plus, Pro, Enterprise or Edu (the Business plan is not on the list)
• A person checks and approves every invoice before it is posted to the software
• A sample of invoices is ready, including Hebrew-language and scanned ones
• The terms of service have been checked for how uploaded supplier data is handled
• The test metrics are set: completion rate, time, cost and review effort
• The agent gets no wider access than the task requires

The home page has screenshots of systems I built. If you have a similar process that takes up time, you can tell me about it in a short conversation.

Frequently asked questions

Does GPT-6.1 Terra work with Priority or SAP Business One?

No source mentions Israeli software or any specific accounting system. According to TechRepublic, the model operates software through its normal screens, so no ready-made integration is needed, but only a trial on the business's own system will show whether it copes. The first trial is best run as a dry run, with a person checking a list and nothing posted.

How does trying it in ChatGPT Work differ from using the API?

ChatGPT Work is manual, one conversation at a time, within the subscription's usage allowance. The API is for automation at volume: billing is per token, and a developer has to build the process. OpenAI's invoice example refers to the API, and that is where the price cut applies.

What if the model misreads an invoice?

That is the point of starting with a review queue: a person corrects the error before anything is posted. Logging each correction is worthwhile, because the number of fixes and the time they take are part of the real cost to compare with the current process.