GPT-6.1 Terra: computer use, half the API price, and what is behind the numbers

On October 26, 2026, OpenAI introduced GPT-6.1 Terra. By description, it is a model focused on two things: computer use (using a computer — when the model itself clicks, enters text, and switches between windows) and detailed visual reasoning. Plus clearer text. Prices in the standard API are lower than those of GPT-5.6 Terra and GPT-6.1 Sol.

What “precise computer use” means

The wording in the announcement is “precise computer use.” The difference between “the model understands a screenshot” and “the model does the work” lies in the cost of an error. If the task is to fill out a form with ten fields, one typo in the account number breaks the entire process. Terra is presented as a step forward relative to GPT-5.6 Terra specifically on communication and computer tasks, which make up the workday.

An example from the announcement: read an incoming invoice, extract the payment details, and enter them into an accounting program. The work is boring and consists of a chain of small actions: open the attachment, find the number, switch to the accounting window, enter, check. It is precisely on such chains that promises and reality usually diverge — therefore the numbers below should be read as the vendor’s claim.

The third component is “clearer and more natural text.” It sounds like marketing, but it has a verifiable consequence: fewer edits after generation. In scenarios such as a draft commercial proposal, edits are the main time cost, not generation itself. How much better the text has become is not measured separately: there are no numbers on writing quality in the announcement.

Half the price: per token or per task?

Standard API prices (the programmatic interface through which the model is called from your code) for GPT-6.1 Terra are 50% lower than those of GPT-5.6 Terra and GPT-6.1 Sol. A token (a unit of text, roughly a piece of a word) is the basic billing unit, so “half the price” means: the same flow of tasks costs roughly half. For a developer, this is the difference between automation “as a test” and automation for the entire incoming document flow.

But price per token is not the same as price per task. A model that spends three times fewer tokens turns out to be cheaper not by half, but by several times. In the announcement this is shown in comparison with Claude Opus 5.5 at maximum effort: more than three times more token-efficient at a quarter of the price per token — the vendor speaks of a ninefold reduction in cost per task, although multiplication gives about twelve. This is the vendor’s arithmetic, not an independent check.

The wording “cost-intelligence frontier” means the price/quality boundary across the model line. The practical meaning: Terra is positioned not as the strongest model, but as the most cost-effective on mass tasks. If the task is one-off and complex, cheapness is secondary. If it is repeated thousands of times a day, the opposite is true.

In ChatGPT Work and Codex, billing is different: there they count not tokens but a usage limit. The announcement’s wording is that the same limit covers more tasks than with GPT-5.6 Terra, and substantially more than with GPT-6.1 Sol or GPT-6 Astra. There are no specific limit numbers — no data.

Agents' Last Exam: what the test is and why it is indicative

Agents' Last Exam is a benchmark (a set of tasks with a fixed evaluation methodology) of professional work scenarios. Two examples from the description: turn an architectural drawing into a 3D model, and review an animated film frame by frame, compiling a log with timecodes for the restoration team. Both require looking at an image, holding context, and producing the result in a structured form.

Results: at medium reasoning effort (reasoning effort — a parameter that sets how much the model “thinks” before answering), Terra is 15 percentage points higher than GPT-5.6 Terra at roughly a third of the cost. On the same test, it reaches the level of GPT-6 Astra at one tenth of the price per token. At maximum effort, it is more than three times more token-efficient than Claude Opus 5.5, at a quarter of the price per token and roughly 9 times cheaper per task. On speed, it is more than three times faster than GPT-6 Astra and GPT-6.1 Sol.

How to read this: 15 percentage points is not “15% better,” but the difference in the share of solved tasks. If the previous model passed 60 out of 100, the new one passes 75 out of 100. Absolute values are not given in the announcement, so it cannot be recalculated — no data.

MMMU-Pro and the limits of the claim

MMMU-Pro is a benchmark of visual reasoning: tasks on understanding diagrams, charts, and images together with text. In the provided fragment of the announcement, the phrase breaks off at “outperforms GPT-5.6 Terra at every reasoning,” so we will not assert specific effort levels and numbers. Only the direction is known — superiority over GPT-5.6 Terra. There are no numbers.

Separately about choosing between models. In the announcement, Terra is cheaper than GPT-6.1 Sol and covers more tasks on the same limit — that is, for mass scenarios the vendor offers precisely it. There is no comparison of quality on common tasks between them in the text, so the boundary will have to be found on your own.

What is not in the announcement at all: timing and access channels, context window size, supported languages, failure statistics, results of independent runs. All the figures given are the vendor’s measurement under the vendor’s conditions. For an announcement this is normal, but before deployment you need your own test on your own data.

And one more thing that is easy to confuse when reading tables: “three times faster” and “three times more token-efficient” are different things. The first is about time, the second about money.

What to do with this

• Build your own set of 20–30 typical tasks — invoices, forms, drawings, draft letters — and run it on GPT-5.6 Terra and GPT-6.1 Terra. Compare the share of tasks accepted without edits and the cost of one task.
• Calculate price per task, not per token: multiply token consumption by price, add retries after errors and manual checking.
• For batch document processing, a 50% reduction in API price is a direct argument: where the flow is large, the difference is visible in the monthly bill.
• In scenarios with clicks and forms, leave human checking at the last step: there are no error statistics in the announcement, and the cost of one incorrect entry is higher than the cost of tokens.
• If you work through ChatGPT Work or Codex, recalculate the limit for your tasks: it is claimed that the same limit covers more, but no specific numbers are given.