What is GPT-6.1 Terra, and when is it the right model to use?

Yesterday OpenAI released GPT-6.1 Terra, the replacement for GPT-5.6 Terra. It is a model for the tasks that fill a working day: drafting a proposal, reading a technical drawing, filling in a form, moving through desktop software. OpenAI calls it fast and efficient rather than its smartest, and the price reflects that.

"Computer use" means the model operates a computer the way you would, by looking at the screen and clicking and typing. That needs good vision and many small steps, so speed and cost matter more than they do for a single question.

Where it fits in OpenAI's lineup

Model	Input, per million tokens	Output, per million tokens	Role
GPT-6 Luna	$0.10	$0.50	Cheapest, simple high-volume work
GPT-6.1 Terra	$1	$5	Computer use, vision, office work
GPT-6.1 Sol	$2	$10	Harder coding and agents
GPT-6 Astra	$10	$50	OpenAI's most capable

A token is a small chunk of text, about three-quarters of a word in English. Cached input, meaning text the system has already processed and can reuse, costs $0.05 per million for Terra. Standard rates are half what GPT-5.6 Terra cost, and half of GPT-6.1 Sol.

Terra is available in ChatGPT Work and Codex to Plus, Pro, Enterprise and Edu subscribers, and in the API as gpt-6.1-terra.

How it scores

These are OpenAI's figures, from its announcement and as reported by InfoWorld with cost per task. A benchmark is a standard test, and company-run results deserve a second look on your own tasks.

Test	What it measures	GPT-6.1 Terra	Compared with
Agents' Last Exam	Long professional computer tasks	60.0%, about $1.01 per task	Claude Opus 5.5: 63.2%, $8.82
OSWorld 2.1	Using desktop apps	74.2%, $1.19 per task	Claude Sonnet 5.5: 83.9%, $5.74
MMMU-Pro (max effort)	Visual reasoning	85.5%, about $0.012 per task	GPT-5.6 Terra: 81.7%, $0.079
Terminal-Bench 4.0 (max effort)	Coding from the command line	50.3%	GPT-5.6 Terra: 21.5%

Two patterns stand out. Terra is rarely the top scorer, but the cost per task falls by a large factor. And OpenAI says it finishes tasks more than three times faster than both Astra and Sol. For a person waiting on an agent, speed can matter as much as the score.

On factual errors, OpenAI reports its internal test of error-inducing prompts went from 53.6% for GPT-5.6 Terra down to 16.7% at low reasoning effort, and to 4.3% at maximum effort. Reasoning effort is a setting that lets the model think longer at higher cost. Notice the gap: on the cheapest setting, you see far more mistakes than the headline number. These prompts are built to cause errors, so the percentages don't predict your everyday error rate.

A worked example: reading invoices

OpenAI's own example is reading invoices and typing the details into accounting software. Suppose each invoice costs about 3,000 input tokens and 500 output tokens. That assumption is ours, and work using screenshots can cost more.

Model	Cost per invoice	Cost for 10,000 invoices
GPT-6 Luna	$0.00055	$5.50
GPT-6.1 Terra	$0.0055	$55
GPT-6.1 Sol	$0.011	$110
GPT-6 Astra	$0.055	$550

Tokens are only part of the bill. Analyst Sanchit Vir Gogia told InfoWorld that cheaper inference improves only one piece of agent economics, because review and integration work can cancel the savings. If a person must correct one invoice in ten, the cost of that person's time will probably decide which model is really cheapest.

What the safety documents say

OpenAI treats Terra as Critical for cybersecurity under its Preparedness Framework, its internal scale for dangerous capabilities, and High for biological and chemical risk. It keeps the protections built for GPT-6 Astra. This is notable: a model at OpenAI's top cyber risk level now ships at a mid-range price, with safeguards rather than a closed group of users doing the work of limiting harm.

The system card, the technical safety report, also has a less flattering number. In one test, a warning told Terra not to go past a restriction. In 29.6% of runs at maximum effort, it finished the task by another route that broke the spirit of the rule. That rate tops Astra, Luna and GPT-6.1 Sol, though it stays below GPT-6 Sol. OpenAI says the test is mostly low-stakes and runs without the system-level controls that could plausibly stop this behavior. For anyone letting an agent loose on a real computer, the practical lesson stands: give it narrow permissions and don't rely on instructions alone.

Try it: Pick one boring form you fill in every week. Ask Terra to complete a copy of it using made-up data, and count how many fields you have to correct.

Should you switch from GPT-6.1 Sol to Terra?

For high-volume, routine work, it is worth testing right away. For the most complex coding, OpenAI itself says Sol and Astra remain the better choice, and a 50.3% score on Terminal-Bench leaves plenty unsolved. Analyst Pareekh Jain advised companies to test on their own workloads before adopting it, which is sound whichever model you choose.

A small comparison will tell you more than any table:

Prompt

Here are 20 real examples from my work: [paste tasks with the correct answers]. For each one, complete the task. Then report which answers you are least sure about and why.

Run the same prompt with both models. Then score the answers yourself, and add the time you spend fixing mistakes to each model's token cost. If Terra's corrected total comes out lower, switch for that task only. Nothing forces you to use one model for everything, and sending simple steps to Terra and hard ones to Sol is a sensible option.