# GPT-6.1 Terra arrives at half the price and with more PC use

> GPT-6.1 Terra charges 1 dollar per million input tokens and 5 for output, half that of GPT-6.1 Sol, and uses the computer on par with GPT-6 Astra.
> By José Maldonado · Artificial Intelligence · 2026-10-28

GPT-6.1 Terra is now out. OpenAI presented it this Monday, October 26, with a rate of 1 dollar per million input tokens, 0.05 per million cached input and 5 per million output, and opened it the same day to Plus, Pro, Enterprise and Edu subscribers in ChatGPT Work and Codex, as well as the API under the identifier \`gpt-6.1-terra\`, as reported by [the model presentation](https://openai.com/index/introducing-gpt-6-1-terra/). Exactly half what GPT-6.1 Sol charges per token.

What stands out is not the price cut, but what the cheap model does: reading an architectural blueprint and returning a three-dimensional model, reviewing a film frame by frame to leave a record with timecode, filling in forms on screen or matching product photos with their catalog listing. That is the work that until now forced you to pay the top-of-the-range model's rate, and OpenAI now puts it at the level of GPT-6 Astra for a tenth of the price per token.

Therein lies the shift. When the task is billed per run and not per token, the price list stops deciding on its own and what matters is how much the model consumes to solve it, so a mid-range model can end up costing more than the flagship if it reasons three times as much. OpenAI sells four classes of model and places the cost-per-task argument precisely on the one that carries a company's repetitive work.

## A third of the cost for the same task

In Agents' Last Exam, OpenAI's professional-task benchmark, GPT-6.1 Terra at medium effort scores 15 percentage points more than GPT-5.6 Terra at a third of the cost. It also performs at the level of GPT-6 Astra while paying a tenth per token and finishes each task more than three times faster than Astra and Sol.

It is worth qualifying where those figures come from: they are published by the lab that sells the model, on a benchmark it designed itself, and in the comparison with Claude Opus 5.5 it does not detail version or parameters. They serve to frame the conversation, not to close a purchase: the test that decides is your workflow, with your tools and your bill.

The exam tasks are those that until recently required a person in front of the screen. Converting a blueprint into a three-dimensional model means reading symbols, holding the geometry and not losing dimensions, and reviewing an animation frame by frame to note every issue is a restorer's work.

## The table that explains the shift

Model

Input per million

Cached input

Output per million

GPT-6.1 Terra

1,00 $

0,05 $

5,00 $

GPT-6.1 Sol

2,00 $

0,10 $

10,00 $

GPT-5.6 Terra

2,00 $

0,20 $

12,00 $

GPT-6 Astra

10,00 $

1,00 $

50,00 $

GPT-6 Luna

0,10 $

0,01 $

0,50 $

These are the standard short-context rates and come from the official API list, where cached input is 5% of the normal rate, according to [the full table](https://developers.openai.com/api/docs/pricing). Batch and Flex mode cut another 50%, fast mode doubles the price and regional processing adds 10%.

The cut is best measured against the model it replaces. Compared with GPT-5.6 Terra, input falls by half and output drops from 12 to 5 dollars, a cut of more than half, while against GPT-6.1 Sol the reduction is exactly 50% across all three items.

A detail the table hides: above 272 thousand input tokens, the entire request costs double for input and one and a half times for output. Cache writing costs 1.25 dollars with 30 minutes of retention and 2.50 with 12 hours, so the saving depends on the context being repeated.

## Fewer errors at low effort

OpenAI also publishes an internal benchmark of instructions that induce errors. At low effort the rate falls from 53.6% for GPT-5.6 Terra to 16.7% for GPT-6.1 Terra, almost 37 percentage points. At maximum effort it drops another 18 points to 4.3%, practically on par with Astra and Sol.

The model accepts six levels, from \`none\` to \`max\`, with \`medium\` as the default value. Its spec sheet declares 1.05 million tokens of context, 922 thousand maximum input and 128 thousand output, according to the [model technical sheet](https://developers.openai.com/api/docs/models/gpt-6.1-terra). The knowledge cutoff remains June 1, 2026 and input accepts text and image.

The edges are just as clear: there is no audio or video, neither input nor output, and the model only returns text, so an agent that needs to speak or transcribe will have to chain another piece. Tool calls require the Responses API.

## The plan goes further and there is the fine print

The other half of the announcement is aimed at those who pay a subscription and do not bill by API. With the same usage allowance, Terra covers more tasks than GPT-5.6 Terra and considerably more than GPT-6.1 Sol or GPT-6 Astra, according to the launch note. What is new is that computer work no longer uses up the expensive model's quota.

The fine print is that OpenAI does not publish how many tasks fit into each plan or how system tool usage is counted. Anyone managing a team will see the saving on the API bill, but in the subscription they only have a direction of change with which to budget the quarter.

## From price per token to cost per task

OpenAI's argument is that token efficiency weighs more than the rate. At maximum effort, Terra is more than three times more token-efficient than Claude Opus 5.5 and costs a quarter per token, which leaves a cost per task about nine times lower, according to the lab's own comparison.

In Terminal-Bench-Science 0.1, a benchmark of scientific coding tasks, Terra at maximum effort beats GPT-6.1 Sol at lower cost and comes within five points of Opus 5.5 while paying less than a fifth. Hence a useful rule before migrating a workload: the model with the lowest rate that reasons three times as many tokens ends up costing more.

That means the unit of purchase has moved. For two years the conversation was how much a million tokens costs; now it is how much a read invoice or a converted blueprint costs, and that figure does not appear on any price list until the team measures it in its own workflow.

## What the announcement leaves unanswered

What is always missing on launch day: an independent measurement. No external evaluator has yet reproduced the 15 points on Agents' Last Exam or the tie with GPT-6 Astra, and recent history suggests waiting for the verdict of someone who is not paid by OpenAI.

The closest precedent is from July, when OpenAI put the broken tasks on SWE-Bench Pro at close to 30% and withdrew its recommendation to use it, as documented by [the July 8 report](https://openai.com/index/separating-signal-from-noise-coding-evaluations/), just as its model fell behind on that benchmark. A lab that audits benchmarks also chooses which ones it audits.

It also remains to be seen what happens when computer use fails. An agent that fills in forms gets the wrong field, and the cost of that error is not measured in tokens but in the person who has to undo it. That is the ground where a mid-range model stakes its reputation, and no lab benchmark captures it entirely.

## What changes when the cheap model ties with the expensive one

OpenAI had already separated profiles within the same version when it launched [GPT-5.1 Instant and GPT-5.1 Thinking](https://observatorioblockchain.com/inteligencia-artificial/que-trae-el-nuevo-gpt-5-1-de-openai/), and that move was a choice of response style. Terra is a profit-and-loss decision, because the model that ties on the task is precisely the one that costs half.

With four classes of model in the catalog, the decision shifts from the shop window to the accountant. If the middle tier does the flagship's work at a tenth of the price per token, the question stops being which model is better and becomes how many runs fit into the month's budget.

That cheapening has an uncomfortable side. The same cost per task that makes automating invoice reading profitable also serves to monitor from a distance, and OpenAI is already under judicial pressure over how it handles minors' data, as reported by [the lawsuit from eight states](https://observatorioblockchain.com/en/ai/ocho-estados-demandan-openai-proteccion-menores/). The cost per task falls while the legal bill rises.
