GPT-6.1 Terra Review: Computer Use for About $1 a Task (We Tested It on 20 Invoices)

AI that can use a computer has been expensive. GPT-6.1 Terra, out October 26, 2026, changes the price. On OpenAI's own test of hard computer work, it scores within about three points of Claude Opus 5.5, for one dollar a task instead of $8.82.

We went through every chart OpenAI published, number by number, then ran our own test: 20 trap-filled invoices through 5 setups. All five got every field right. Terra did it for about a tenth of what Opus cost.

The demo: a tram becomes a toy

OpenAI's launch video shows an AI one photo of a Lisbon tram. It opens a 3D app by itself, clicks through every tool, and hands back a model a toy factory could build. The clip is shown at real speed, with some sections cut. Nice demo. But the part that matters is the bill.

What Terra is, and what it costs

Terra is the middle of OpenAI's lineup. Astra is the big one, Sol is the flagship, Luna is the cheap one, and Terra sits between them.

• Price: $1 per million tokens in, $5 out. (A token is a small chunk of text, roughly three quarters of a word.) That's half of Sol and a tenth of Astra.
• What it handles: text and images, about a million tokens at once, and computer use: clicking, typing, filling in forms.
• Where: ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu, and the API as gpt-6.1-terra.

Two bits of fine print. The launch video says "all paid plans," but the article only names those four. If you're on Go or Business, check your model menu. And past about 272 thousand tokens in one request, the input price doubles.

Computer use: Agents' Last Exam

Agents' Last Exam is a set of real office jobs. Turn architectural drawings into a 3D model. Go through an animated film shot by shot and write a time-coded log for a restoration team.

At medium effort (how hard the model thinks before answering), the new Terra gets 57.6%. The old Terra got 42.6%, and it cost almost three times as much per task. Astra, OpenAI's biggest model, gets exactly the same 57.6% at medium. It costs $4.82 a task, against Terra's 29 cents.

Turn Terra all the way up and it reaches 60% for about a dollar. That's above Sol and Astra at their max. Claude Opus 5.5 still tops the chart at 63.2%, but each task costs $8.82. OpenAI also says Terra finishes these jobs more than three times faster than Astra and Sol. The chart doesn't show time, so that one is their claim.

Desktop apps: OSWorld 2.1

OSWorld tests whether an agent can work real desktop apps through jobs with many steps. Here's the headline: at medium effort, new Terra scores 66.7% for 50 cents a task. The old Terra, pushed to its max, only reached 57.1%, and that cost $4.76. A higher score for about a tenth of the money.

Now look at the top of OpenAI's own chart. Sol at max gets 78% for almost the same price as Terra at max ($1.15 vs $1.19). Claude Sonnet 5.5 gets 81% at $2.22, and 83.9% at max. So on desktop apps, Terra is the cheap way in. It's not the best score on this chart.

Seeing the screen: MMMU-Pro

Before it can click anything, it has to see the screen. MMMU-Pro is a hard test of reading images, like charts, diagrams and technical drawings. Terra at its lowest effort scores 83.7%, and each question costs less than half a cent. Astra gets 85.3%, about a point and a half higher, for ten times the price. The old Terra couldn't reach 83.7% even at max.

Fewer made-up facts

This is the change we care about most. OpenAI has a set of prompts built to trick models into getting facts wrong. On those, the old Terra at low effort had an error in more than half its answers, 53.6%. New Terra is at 16.7%. At max effort it drops from 22.3% to 4.3%, basically the same as Sol and Astra.

These are trick prompts, so your everyday error rate should be lower. But if Terra is filling in your forms, this is the number that decides whether you have to check every field.

Coding: Terminal-Bench 4.0

Terminal-Bench 4.0 gives an agent real problems to fix from the command line, like a webpage whose layout jumps while it loads. Terra at max solves 50.3%, more than double the old Terra's 21.5%, for about a third of the price.

But read the chart carefully. Sol at medium scores higher, 52.7%, and it's cheaper: $1.02 a task against Terra's $1.84. OpenAI says it in the article too: for the hardest coding, use Sol or Astra. Claude Sonnet 5.5 at max reaches 70.6%, at $10.44 a task. If you code in Codex, keep Sol for the hard stuff.

Science: Terminal-Bench-Science

This chart flips the coding result. Terminal-Bench-Science gives agents real research jobs in code, like tracking cells dividing in a time-lapse video. Here Terra at max beats Sol, 58.3% against 57%, and it's cheaper. Claude Opus 5.5 scores 63.3%, but at $23.21 a task. Terra is five points behind for under a fifth of the price. Astra has the top score at 68.1%, at $23.80.

Safety: the system card

A cheap model that can use a computer is also a cheap model for attackers. OpenAI rates Terra Critical for cybersecurity, the top level in its own Preparedness Framework. In their test on recently disclosed security holes, Terra wrote a working exploit 24.5% of the time. That's more than Sol.

So it ships with the same safeguards as Astra. OpenAI also says it resists prompt injection better. That's the trick where a webpage hides instructions to take over the agent. Any agent that reads web pages should still get the fewest permissions it needs.

Our test: 20 invoices, 5 models

OpenAI's own example for Terra is reading invoices and typing them into accounting software. So we made 20 fake invoices and filled them with traps:

• A deposit that's already paid, and "Net 30" instead of a due date
• An order number sitting right next to the invoice number
• German and French number formats, UK prices that already include VAT, Japanese yen with no decimals
• A handwritten "paid $200 cash" note and a PAID stamp
• A dollar sign that's actually Canadian dollars
• A credit and a late fee on one bill, two tax lines to add up
• Three invoices shot crooked on a phone

Every model had to fill in the same nine fields: vendor, invoice number, invoice date, due date, currency, subtotal, tax, total and amount still due.

All five setups got all 180 fields right. Every trap, every model. So the only difference is the bill. Ten thousand invoices would come to $11.85 with Terra, about $65 with Sol or Sonnet, and $133.42 with Opus. That's eleven times what Terra cost, for exactly the same answers. Turning Terra up to max made it slower and three times pricier, and it got nothing more right.

To be fair

Two limits. This test reads the invoice; it doesn't click through an accounting app. And when everyone gets everything right, the test can't tell you which model is smarter. It only tells you that for this kind of job, paying more bought us nothing.

Which model for which job

• Screen work, forms, invoices (anything where the AI reads a page and clicks): start with Terra at medium effort. It's the best value on every computer-use chart OpenAI published.
• Hard coding: stay on Sol.
• Tiny jobs in bulk: Luna, at $0.10 in and $0.50 out per million tokens.
• Only the top score matters: the most expensive models still win, and that includes Claude.

Sources: OpenAI's launch post, the gpt-6.1-terra model docs and the GPT-6.1 Terra system card addendum.

Get the Terra Fare Card

We put every number from the video on one page: all six OpenAI charts as tables, our invoice results and the routing rule, so you don't have to pause and screenshot. Comment TERRA on the video, or grab it below.

How we tested: 20 fake invoices (made-up companies), the same 9-field form for every model, run October 26, 2026, one run each, real billed cost. One small test, not a study. Benchmark numbers are OpenAI's own, read from its launch charts. Independent, not affiliated with OpenAI or Anthropic.