GPT-6.1 Terra Just Made Computer Use 9× Cheaper

Author: Hyperautomation Labs · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: GPT-6.1 Terra Just Made Computer Use 9× Cheaper · Channel: Hyperautomation Labs · 07:46 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, Model inference and optimization, AI safety

Summary

The video reviews OpenAI’s new Terra model, focusing on its computer-use abilities and lower per-task costs compared with larger models. Across OpenAI’s benchmarks, Terra offers strong value for desktop tasks, image understanding, and several other workflows, though it does not lead every chart. The presenter also tests five model setups on 20 invoices and finds that all correctly extract every field, while Terra costs much less. The recommended approach is to start with Terra for screen-based work, use Soul for difficult coding, and reserve more expensive models for cases where the highest score matters.

Key points

Terra is positioned between OpenAI’s smaller Luna and larger Saul and Astra models, with pricing of $1 per million input tokens and $5 per million output tokens.
On Agent’s Last Exam, Terra scores 57.6% at medium effort and 60% at maximum effort; the video reports a task cost of about $1 at maximum effort.
On OS World, Terra scores 66.7% at medium effort for about 50 cents per task, outperforming the older Terra at maximum effort for a much lower cost.
Terra performs strongly on image understanding, scoring 83.7% at its lowest effort on MMMU Pro for less than half a cent per question.
OpenAI’s prompt tests show a substantial reduction in Terra’s error rate compared with the previous version, including a drop to 4.3% at maximum effort.
For the hardest coding tasks, the presenter recommends Soul or Astra; Terra performs better on the Terminal Bench Science test than Soul, at lower cost.
OpenAI rates Terra critical for cybersecurity and says it has safeguards like Astra; the video also emphasizes limiting agents to the permissions they need.
In the presenter’s invoice test, all five setups got all 180 fields right, while Terra at medium effort cost about a tenth of a cent per invoice.
The presenter’s usage recommendation is Terra for screen work and forms, Soul for hard coding, Luna for tiny jobs in bulk, and the most capable model when top scores matter.

Chapters

Start	Chapter	Summary
00:01	Terra’s pricing and position	The presenter introduces Terra’s computer-use capabilities, pricing, place in OpenAI’s lineup, and availability.
01:33	Benchmark results across work tasks	The video compares Terra with other models on office tasks, desktop use, image understanding, reliability, and coding.
05:40	Security and the invoice test	The presenter discusses Terra’s security rating and reports that all tested setups extracted every field from the sample invoices.
07:15	How to choose a model	The presenter gives model-selection recommendations and points viewers to a page collecting the video’s numbers and routing rule.

Quotes

But the part that matters is the bill. — 00:01

So, that’s a higher score for about a tenth of the money. — 02:34

this is the number that decides whether you have to check every field. — 04:06

paying more bought me nothing. — 06:43

For screen work, forms, invoices, anything where the AI reads a page and clicks, start with Terra at medium effort. — 07:15

Concepts and entities

GPT-6.1, Terra, Soul, Astra, Luna, Claude Opus 5.5, Claude Sonnet 5.5, OpenAI, OSWorld, MMMU-Pro, Terminal-Bench 4.0, Terminal-Bench Science, prompt injection

Full transcript (original language)

00:01 You show an AI one photo of a tram. It opens a 3D app by itself, clicks through every tool, and hands back a model a toy factory could build. That’s real speed from OpenAI’s launch video for GPT 6.1 data out today. But the part that matters is the bill. On OpenAI’s own test of hard computer work, data scores within about three points of Claude Opus 5.5 for $1 a task instead of $8.82.

00:32 So I went through every chart they published number by number and ran my own test to see where that holds up and where it doesn’t. Data is the middle of OpenAI’s lineup. Astra is the big one. Saul is the flagship. Luna is the cheap one, and data sits between them. The new version costs $1 per million tokens in and $5 out. That’s half of Saul and a tenth of Astra. It reads texts and images. It can hold

01:03 about a million tokens at once, and it can use a computer clicking, typing, filling in forms. Starting today, it’s in ChatGPT work and codex for plus, pro, enterprise, and edu and in the API. Two bits of fine print. The launch video says all paid plans, but the article only names those four. So if you’re on go or business, check your model menu. And past about 272,000 tokens in one request, the input price doubles. Agent’s last exam is a set of real

01:33 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, the new data gets 57.6%. The old data 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.

02:04 Turn Terra all the way up and it reaches 60% for about a dollar. That’s above Soul 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 Soul. The chart doesn’t show time, so that one is their claim. OS World tests whether an agent can work real desktop apps through jobs with many steps. Here’s the headline. At medium

02:34 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. So, that’s a higher score for about a tenth of the money. Now, look at the top of OpenAI’s own chart. Soul at max gets 78 for almost the same price as Terra at max. And Claude’s on it, 5.5, gets 81% at 2.22 and nearly 84

03:04 at max. So, on desktop apps, Terra is the cheap way in. It’s not the best score on this chart. Before it can click anything, it has to see the screen. MMM U 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 10 times the price. The old Terra couldn’t reach 83.7 even

03:35 at max. At the top end, all three new models sit within about one point of each other. This is the change I 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 Soul and Astra. These are hard trick prompts. So,

04:06 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. For coding, Terminal Bench 4.0 gives an agent real problems to fix from the command line. Like a web page whose layout jumps while it loads. Terra at max solves 50.3%. That’s more than double the old Terra’s 21.5 for about a third of the price. But read the chart carefully. Soul at medium effort scores higher, 52.7, and

04:37 it’s cheaper. A dollar two a task against Terra’s dollar 84. OpenAI says it in the article, too. For the hardest coding, use Soul or Astra. And Claude Sonnet 5.5 at max reaches 70.6 at $10 a task. So, if you code in CodeX, keep Soul for the hard stuff. One more chart because it 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 Soul, 58.3 against 57, and it’s

05:10 cheaper. Claude Opus 5.5 scores 63.3, but at $23 a task, Terra is five points behind for under a fifth of the price. Astra is the top scorer at 68, also about $24. 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

05:40 working exploit 24.5% of the time. That’s more than Soul. So, it ships with the same safeguards as Astra. They also say it resists prompt injection better. That’s the trick where a web page hides instructions to take over the agent. Any agent that reads web pages should still get the fewest permissions it needs. OpenAI’s own example for Terra is reading invoices and typing them into accounting software. So, I made 20 fake invoices and filled them with traps. A deposit that’s already paid, net 30 instead of a due date, an order

06:12 number sitting right next to the invoice number, German and French number formats, a handwritten note saying paid 200 cash, a paid stamp, a dollar sign that’s actually Canadian dollars, and three of them shot crooked on a phone. Every model had to fill in the same nine fields. Here’s the result. All five setups got all 180 fields right. Every trap, every model. So, the only difference is the bill. Terra at medium cost, about a tenth of a

06:43 cent per invoice. 10,000 invoices would come to about $12. Soul and Sonnet come to about 65. Opus comes to 133. That’s 11 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. Two limits, to be fair. 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

07:15 job, paying more bought me nothing. So, here’s how I’d use it. For 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. For hard coding, stay on Soul. For tiny jobs in bulk, Luna. And when only the top score matters, the most expensive model still win, and that includes Claude. I put every number from this video on one page with the routing

07:46 rule, so you don’t have to pause and screenshot. Comment Terra, spelled t e r r a, and I’ll send it to you. Follow along on Instagram and Facebook for bite-size AI, and the deep dive playbooks live here on YouTube. Subscribe so you don’t miss one. See you in the next one.