OpenAI's GPT-6.1 Terra clicks through accounting software at $1 per million tokens. Here's the AP test for a sub
OpenAI launched a cheaper computer-use model whose headline demo is entering invoices into accounting software. For a contractor's back office, the question is whether it can key vendor invoices and pay apps without a person re-checking every line.
A subcontractor's accounts payable desk can now test a computer-use AI that is cheap enough to try on real invoices. OpenAI launched GPT-6.1 Terra on October 26, a mid-tier model that works a screen the way a person does: it looks, clicks, types and switches between apps. Its lead demo is reading incoming invoices and entering them into accounting software.
What did OpenAI actually release?
Terra sits below OpenAI's top models and is priced at $1 per million input tokens and $5 per million output tokens, according to coverage from The Next Web and BusinessToday. It is available in ChatGPT Plus, Pro, Enterprise and Edu, and in Codex.
OpenAI's performance claims are its own: roughly the capability of its larger Astra model at about a tenth of the price per token, and much faster completion times. The Neuron's write-up notes independent tests found a smaller speed gap, and that on a 15-field form test run three times, Terra got every field right each time. An OpenAI engineer described it as more a way to stretch usage limits than a breakthrough (The Neuron).
Why should a sub's back office care?
Because most contractor admin is screen work: keying vendor invoices, matching them to POs, entering pay application lines, updating job-cost codes. Until now, automating that meant fragile integrations or a human typing. A model that operates the software directly doesn't need an integration, which is useful for the small shops running older accounting packages.
The cost point matters too. Cheap per-token pricing is what makes a "try it on every invoice" experiment affordable instead of a budget request.
What are the limits?
- Money moves need a person. A wrong cost code or a duplicate payment is a margin problem, not a typo.
- Benchmarks are the vendor's. Nobody has tested it on a retainage-heavy AIA pay app or a lien waiver packet yet.
- Screens change. Computer-use agents can break when a vendor portal or accounting screen updates.
- Data exposure. Pointing an agent at your accounting system gives it your vendor and pricing data. Check your plan's data terms first.
How should a sub test it?
| Step | What to do |
|---|---|
| 1. Pick one task | Vendor invoice entry for one job, not the whole AP process |
| 2. Use a sandbox | A test company file or copied data, never live books |
| 3. Run 50 real invoices | Include messy ones: handwritten, multi-page, credit memos |
| 4. Score against staff | Compare fields keyed: vendor, amount, job, cost code, date |
| 5. Time and cost it | Include review time, not just model time |
| 6. Decide | Keep a human approval step on anything that pays |
The takeaway
Don't buy a back-office overhaul on launch-day claims. Spend two weeks running Terra against 50 of your own invoices in a sandbox, count the field errors, and compare review time to keying time. If the error rate is low and a clerk can approve in seconds, you have a real hours-saved number. If not, you've spent a few dollars finding out.
- What is GPT-6.1 Terra?
- GPT-6.1 Terra is a mid-tier OpenAI model launched October 26, 2026, built for computer use: it looks at a screen and clicks, types and moves between apps. OpenAI's headline example is reading incoming invoices and entering them into accounting software.
- How much does GPT-6.1 Terra cost?
- OpenAI prices it at $1 per million input tokens and $5 per million output tokens through the API. It also appears in the model picker for ChatGPT Plus, Pro, Enterprise and Edu plans, and in Codex.
- Can an AI model enter invoices into construction accounting software?
- Models like Terra can operate an accounting screen directly, but the speed and cost figures come from OpenAI's own testing. A contractor should run its own sample of real vendor invoices and compare the results to what staff keyed.
- Should a subcontractor switch its back office to GPT-6.1 Terra now?
- Not wholesale. Run a two-week side-by-side test on one narrow task, such as vendor invoice entry, and keep a person approving anything that moves money.