GPT-6.1 Terra: AI That Uses the Computer at Half the Price

Someone on your team opens the supplier's system, copies data from an invoice, pastes it into a spreadsheet, checks it against the purchase order, and fills out a form. They repeat this dozens of times a day. On October 26, 2026, OpenAI released a model built for exactly this kind of work: GPT-6.1 Terra (API ID: gpt-6.1-terra), which understands what is on the screen and acts inside applications.

What stands out for anyone in charge of a budget is the price. Let's go through the numbers and what they mean for operations.

What GPT-6.1 Terra does

The title of the official announcement sums up the pitch: "GPT-6.1 Terra: delivering fast and capable computer use at a lower price."

In practice, according to OpenAI and the press that covered the launch, the model understands what appears on screen, acts in applications, fills out forms, reads technical drawings, and drafts proposals. It also edits video and creates 3D models from drawings. OpenAI describes it as a fast, efficient model with clearer writing than its predecessor.

It is a mid-sized model. In the current family, according to InfoWorld, GPT-6 Luna is the cheapest, GPT-6.1 Terra sits in the middle, and GPT-6.1 Sol and GPT-6 Astra are the most capable. Terra replaces GPT-5.6 Terra.

How much it costs

Through the API, the standard price is $1 per million input tokens and $5 per million output tokens. A token is the piece of text the model reads or writes, and it is the unit in which usage is billed. With caching, input drops to $0.05.

For comparison, GPT-5.6 Terra cost $2 for input and $12 for output. Input fell 50%. Output fell about 58%, according to the price table. GPT-6.1 Sol is at $2 for input and $10 for output. One outlet converted the values to R$ 5.50 and R$ 27.50, but that is an approximate conversion, not an official price.

There are two more cost rules, according to datanorth.ai: Batch and Flex modes cost half, and requests with more than 272,000 input tokens pay double on input and 1.5 times on output.

For your company, the final cost depends on volume and on the size of each task. A figure in local currency only makes sense after you look at your own case.

What OpenAI says about performance

All the numbers below come from OpenAI itself. The sources consulted do not include independent tests.

• On Agents' Last Exam, with medium reasoning, Terra comes in 15 percentage points above GPT-5.6 Terra, at about one third of the cost per task.
• OpenAI says it matches GPT-6 Astra on that same test, at one tenth of the price per token.
• It completes tasks, on average, more than 3 times faster than GPT-6 Astra and GPT-6.1 Sol.
• On Terminal-Bench 4.0, it scores 50.3% with maximum reasoning, versus 21.5% for the previous Terra.
• In an internal test with questions designed to induce errors, the rate of factual errors fell from 53.6% on GPT-5.6 Terra to 16.7% with low reasoning and 4.3% with maximum reasoning.
• Against Anthropic's Claude Opus 5.5, at maximum reasoning, OpenAI claims Terra uses more than 3 times fewer tokens and costs about 9 times less per task. The competitor's price does not appear in the material.

The error figure deserves attention from anyone thinking about automating verification: even in the best scenario, 4.3% is not zero. Any workflow involving money needs human review of the output.

Where this fits into the routine of a midsize company

Three tasks match what the model does.

Checking spreadsheets and documents

Comparing a purchase order, an invoice, and an entry in the system is work for the eyes and the mouse. A model that sees the screen and fills out forms can do the first pass and leave only the discrepancies for the team. The hour that comes back is that of the person who today checks line by line.

Triaging tickets and orders

Reading the request, classifying it, and opening the record in the right system is pure repetition. The gain is in answering the customer sooner and freeing up the people who were only forwarding requests.

Proposals and reading drawings

For companies that quote based on projects, the ability to read technical drawings and draft proposals shortens the time between request and response.

How to start without betting the budget

My recommendation, and this is my own analysis, not source data: pick a single workflow, high in volume and low in risk, such as spreadsheet checking or ticket triage. Measure how much time and money each task consumes today. Run the model on a sample, with a person reviewing the result, and compare cost and time per task. If the numbers add up, scale up. If they don't, you spent little to find out.

One warning that applies to any project: AI doesn't fix a messy process. If today's checking depends on rules that only one person knows, the first job is to write those rules down.

Where it is available and what is still unknown

Since October 26, the model has been rolling out to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, as well as the API. According to briefia.fr, it is not yet in public ChatGPT, with arrival expected later. DataCamp notes that the announcement does not mention the free plan and that the rollout may take a while to reach all accounts. No source consulted confirms its presence in GitHub Copilot.

On usage limits in ChatGPT Work and Codex, OpenAI says Terra goes further than GPT-5.6 Terra, GPT-6.1 Sol, and GPT-6 Astra.

There are also two limitations. OpenAI explains that GPT-6.1 Sol and GPT-6 Astra are better suited to the most complex programming tasks. And briefia.fr, the only source on this point, reports that Terra reaches the "critical" risk level, the highest on OpenAI's internal cybersecurity scale, and therefore receives the safeguards designed for Astra. Anyone granting access to internal systems should clearly define what the model can and cannot touch.