OpenAI released GPT-6.1 Terra on Sunday — a model built to look at screens, click buttons, and fill out forms at $1 per million input tokens and $5 per million output tokens. That's half the price of its predecessor and roughly a ninth the cost per task of Claude Opus 5.5 at maximum reasoning effort. Within hours, DualEntry ran it through their Accounting AI Benchmark and found it scored just 53.8% on accounts payable — tied for last among 22 models tested.
The story here isn't that Terra is bad. It isn't. It's that we've crossed a cost threshold where AI can do screen-based office work for cents — and the immediate temptation is to hand it your books. The gap between what these models excel at (reading a screen, clicking through a form, classifying a transaction) and what businesses actually need automated (accounts payable, receivables, month-end close) is where real money gets lost.
What Terra actually does
Terra is OpenAI's mid-tier model, positioned below GPT-6.1 Sol and GPT-6 Astra in raw intelligence but designed to be fast, visual, and cheap. It sees screenshots, interprets what's on screen, and takes actions — clicking, typing, scrolling — like a human operator would.
On OpenAI's own benchmarks, the results are impressive. On Agents' Last Exam, a benchmark of professional computer-use tasks, Terra at maximum reasoning effort is over three times as token-efficient as Claude Opus 5.5 and one-fourth its price per token, resulting in approximately 9x lower cost per task. It completes tasks over three times faster than both GPT-6 Astra and GPT-6.1 Sol.
The hallucination rate dropped dramatically too. Artificial Analysis found Terra guesses when it doesn't know the answer 38% of the time, down from 88% for GPT-5.6 Terra — a 50-percentage-point improvement. And on Terminal-Bench 4.0, Terra at max effort solves 50.3% of coding tasks, more than double its predecessor's 21.5%.
The twenty-three cent invoice
InferenceWire broke down the maths: a 20-step invoice entry using around 175,000 input tokens and 10,000 output tokens costs about 23 cents with Terra. When GPT-6 Astra launched in September at $10/$50 per million tokens, that same invoice cost about $2.25 — more than the clerk doing it manually. Terra flips that equation.
This matters most for jobs that never had an integration. As InferenceWire's Abdul Rehman Asghar noted, "Software that was never built to talk to other software — such as an old accounting package, a supplier portal or a clinic's booking screen — can only be automated by something that uses it the way a person does." At 20 cents a task, dozens of small screen-based jobs that were never worth a custom API integration are suddenly worth a trial afternoon.
For businesses still running legacy systems without modern APIs — and that's most small businesses — this is genuinely new territory. We wrote about Xero's XeroForce no-code agent builder earlier this year, which approaches automation from the integration side. Terra approaches it from the screen side. Both have a role, but the screen approach works even when the software vendor doesn't cooperate.
The accounting reality check
This is where the story turns. DualEntry, an independent benchmarking firm, ran Terra through 101 synthetic accounting tasks across eight areas on launch day. The overall score: 78.2%, tied for 18th out of 22 models. Not terrible — but not competitive.
The breakdown tells a sharper story. Terra scored 100% on transaction classification and 92.3% on journal entries — genuinely useful results. But on accounts payable, it managed just 53.8%, tied for last with GPT-6 Luna. Accounts receivable (58.3%) and month-end close (58.3%) were similarly weak.
The consistency numbers are arguably worse. Terra passed only 65.3% of tasks in all three benchmark runs, the second lowest result. For accounting, where the same task runs many times a month, inconsistency means review work.
"GPT-6.1 Terra is one of the fastest models we've tested, and for high-volume work like transaction classification that speed is useful," said Benedict Dohmen, co-founder of DualEntry. "But it missed more simple accounting tasks than almost any other model, which is why finance teams should test AI on the actual work they need done, not on speed or a general benchmark score."
Google's Gemini 4 Argon leads at 91.1%. Three cheaper models — DeepSeek V4.1 Flash, GLM-5.3 Flash, and GLM-5.3 — all scored higher than Terra while costing less per task. This isn't a case where you have to pay more for accuracy; the market already offers it cheaper.
The broader research confirms this isn't a Terra-specific problem. The APEX-Accounting benchmark from Ramp Labs and Mercor found the best model achieved just 56.4% of expert criteria across accounting tasks, with only 21.5% of tasks solved reliably. And ERPBench, which tests agents inside live ERP systems, found that even when agents navigate to the right form and save it, the stored record is often wrong — one model committed saves in 85% of runs but wrote the correct value in just 3%.
What this means for your business
The practical takeaway is straightforward: match the model to the task, not the marketing.
Where Terra earns its place: Transaction classification, data entry into forms, copying information between screens, matching product photos to catalogue entries, interpreting technical drawings. These are tasks where the model's speed and visual reasoning shine, and where a wrong answer is caught by the next step in the workflow.
Where you need something else — or a human: Accounts payable, receivables, month-end close, anything that directly moves money or produces a number someone will rely on. Here, accuracy and consistency matter more than cost, and the benchmark data says Terra isn't competitive.
The sensible architecture, as InferenceWire suggested, costs nothing extra: let the model enter invoices as drafts, and keep two things for people — approving payments and editing bank details. The model does the typing. Humans keep the judgment. If your team has been stung by AI cost blowouts — as Uber discovered when it exhausted its entire 2026 AI budget in four months — Terra's pricing is part of the answer. But cheap and wrong is still wrong.
What to watch
Three things will determine whether computer-use agents like Terra cross from useful-with-supervision to genuinely autonomous in finance work.
First, accuracy on structured tasks needs to break 90% consistently. Gemini 4 Argon is already at 91.1% on DualEntry's benchmark. When the mid-tier models catch up, the cost-accuracy tradeoff disappears.
Second, audit trails for screen-based agents. As The Firm noted, when an agent clicks through your accounting software, every change appears under a staff member's name unless you give it a dedicated account. The governance question is as important as the accuracy question.
Third, prompt injection resistance. Terra resists malicious instructions embedded in documents 98.8% of the time, up from 96.2% for its predecessor. That still means about 1 in 80 hostile invoices could redirect the model. For something with your accounting software open, that's a risk worth managing deliberately.
The cost barrier to AI automation is falling fast — faster than the accuracy barrier. For business owners, the right response isn't to wait, but to start with the tasks where cheap and fast matters more than perfect, and keep humans in the loop where it doesn't.
Sources
- Introducing GPT-6.1 Terra — OpenAI
- GPT-6.1 Terra Is Fast and Cheap. On Accounting, It's Not Accurate Enough to Lead — DualEntry
- OpenAI's GPT-6.1 Terra Is Fast and Cheap but Trails Leading AI Models on Accounting Work — BusinessWire
- GPT-6.1 Terra pushes the cost efficiency frontier — Artificial Analysis
- AI Now Does Office Screen Work for Cents a Task — InferenceWire
- ERPBench: A State-Grounded Evaluation Paradigm for Computer-Use Agents in Enterprise Software — arXiv
- APEX-Accounting — Ramp Labs × Mercor
