GPT-6.1 Terra for invoice entry: screen agent or API?

OpenAI released GPT-6.1 Terra on October 26, 2026, and its launch post names one job directly: reading incoming invoices and entering their details into accounting software. If someone in your office spends hours a week retyping vendor bills, that line is aimed at you.

This post covers what Terra is, how to work out what it would cost for your invoices, and when a model that drives the screen is the wrong tool. The facts about Terra come from OpenAI's GPT-6.1 Terra announcement, which we read on October 27, 2026. What we say about QuickBooks comes from Intuit's help page on emailing receipts and bills.

Terra is OpenAI's fast, lower cost model for computer use, which means it can look at a screen and operate apps: clicking, typing and filling out forms. OpenAI describes it as a big step up from GPT-5.6 Terra.

The parts that matter for back-office work:

• Price. Standard API pricing is $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens. OpenAI says that is 50% lower than GPT-5.6 Terra and GPT-6.1 Sol.
• Where you get it. It is in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API as `gpt-6.1-terra`.
• Fewer wrong answers. On OpenAI's own test of prompts designed to cause errors, the error rate at low reasoning effort fell from 53.6% with GPT-5.6 Terra to 16.7%. At maximum effort it was 4.3%.
• Prompt injection. OpenAI says Terra showed strong resistance to instructions designed to pull it off the user's task.

Those are OpenAI's numbers from its own tests. Your invoices, your vendors and your accounting software are what decide whether it works for you.

Before pricing an agent, it helps to see the options side by side.

1. 1Use the capture your accounting software already has. Intuit's help page says you can forward receipts and bills to a QuickBooks Online address, and QuickBooks pulls out details like dates and amounts so you can review and save the transaction. If that covers your invoices, you may not need anything new.
2. 2Read the invoice with AI, then post it through the accounting software's API. A model reads the PDF and returns the fields, and a workflow writes the bill through an official connection. QuickBooks Online and Xero both offer accounting APIs that cover bills. Our QuickBooks integration and Xero integration pages list what those connections can reach.
3. 3Let a computer-use agent type it in. The model opens the accounting app on a screen and fills in the bill form the way a person would. This is what Terra is built for.

Option 3 is the right pick in fewer cases than the launch post suggests, but there are real ones:

• There is no API. Older desktop accounting software, a supplier portal, or an industry system with no way to connect except the screen.
• The API does not let you do the step. Some systems allow reading by API but only take certain entries through their own screens.
• The volume is low and the format varies a lot. A few dozen odd invoices a month, where building and maintaining an integration costs more than it saves.

We covered this choice in more depth in our post on when browser agents are the wrong surface. In brief, if there is a supported API, use it, and save screen control for the gaps.

• Your software already captures bills. If forwarding to QuickBooks or Xero gets the fields right most of the time, an agent adds cost and a new way to fail.
• You need a clean record of every change. An API call leaves a clear log of what was written. A screen agent clicking through forms is harder to audit.
• Screens change. A software update that moves a button can break a screen agent overnight. An API changes less often and usually with notice.
• Payments are involved. Entering a bill and paying it are different risks. We would never let an agent approve or pay without a person, as we explain in approval gates for finance AI agents.

We cannot tell you what one invoice costs with Terra, because OpenAI does not publish a per-invoice number and it depends on how many screens the agent looks at and how long each run takes. You can measure it in an afternoon:

1. 1Pick 20 real invoices that cover your usual vendors and a few odd ones.
2. 2Run each one through a test setup and record the input and output tokens the API reports.
3. 3Multiply input tokens by $1 per million and output tokens by $5 per million, then average across the 20.
4. 4Add the time a person spends checking each entry, since you should check them all at first.

Compare that total against what the same invoices cost today in staff time, and against what option 1 or 2 would cost. A computer-use run can involve many screenshots per task, while reading a PDF is one pass. That is why the price drop matters, and also why measuring beats guessing.

For a business entering bills by hand into accounting software, here is the setup we would start with:

1. 1One intake inbox. Vendors send invoices to one address. A workflow picks them up and saves each file.
2. 2Read and check. A model pulls vendor, invoice number, date, line items, tax and total. The workflow checks that the lines add up to the total and that the vendor and invoice number are not already in the books.
3. 3Post the easy ones. Known vendors under an amount you choose go in as draft bills through the API, or through Terra if there is no API.
4. 4Send the rest to a person. New vendors, mismatched totals and anything over your limit go to a review queue with the original invoice attached.
5. 5Never pay automatically. Approval and payment stay with a person.

Run it next to your current process for two or three weeks before trusting it, and count how many entries needed fixing. Our post on accounting firm workflow automation covers more of the back-office jobs this fits with.

Terra is new, and the numbers so far are OpenAI's. Watch for independent tests on real accounting apps, how often the agent gets stuck on pop-ups or logins, and whether your accounting software's terms allow automated logins. And because a screen agent reads whatever is in front of it, keep it away from email and websites it does not need, so planted instructions cannot reach it.

Count how many invoices you enter each month and how long each takes. If the answer is a few hours a week or more, it is worth testing. Start with what your accounting software already does, then decide whether an API workflow or a computer-use agent fills the gap.

If you want help with that decision, our free automation audit looks at how invoices reach your books today and tells you which route we would build and why. You can also see what our workflow automation and systems integration work includes.