OpenAI rates GPT-6.1 Terra Critical for cyber. Name who signs what you cannot undo.

On October 26, 2026, OpenAI introduced GPT-6.1 Terra, a model it says combines visual reasoning and computer use, and said its standard API prices are 50% lower than those of GPT-5.6 Terra and GPT-6.1 Sol. The same day, an addendum to the GPT-6 Astra system card said OpenAI is treating GPT-6.1 Terra as Critical in cybersecurity and High for biological and chemical capability. Which tasks still wait for a person is the standing subject of human gates for AI.

This is an illustrative reading of OpenAI’s posts. It adds no measurements of its own, and it does not say whether a team should adopt the model.

What OpenAI reported

OpenAI’s release index and the system-card addendum both date the introduction October 26, 2026.

The product post says GPT-6.1 Terra is available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, and in the API as gpt-6.1-terra. Company-reported standard prices are $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens. OpenAI says those standard prices are 50% lower than the prices for GPT-5.6 Terra and GPT-6.1 Sol.

One use the post names is reading incoming invoices and entering their details into accounting software.

On an internal benchmark of error-inducing prompts, OpenAI reports that the error rate at low reasoning effort fell from 53.6% with GPT-5.6 Terra to 16.7% with GPT-6.1 Terra. At maximum effort, OpenAI reports that the error rate fell by 18 percentage points to 4.3%, and says that figure nearly matches the rates for GPT-6 Astra and GPT-6.1 Sol. Those numbers are company-reported. The prompts are OpenAI’s error-inducing set. They are not an error rate for invoices or any other job of yours.

The product post says the model meets the Critical threshold in cybersecurity under OpenAI’s Preparedness Framework, and that it maintains the protections developed for GPT-6 Astra. The addendum says OpenAI is treating the model as Critical in cybersecurity and High for biological and chemical capability. Based on that assessment, it says, the model uses the same safeguards stack as GPT-6.1 Sol and GPT-6 Astra, with probes updated to be compatible with the model.

What the price does not decide

OpenAI says the lower standard price makes it practical for developers to automate high-volume work. That does not decide which moves still need a person.

The Of One method asks for a named human reviewer before a legal, medical, financial, safety, policy, or reputation move, and before other irreversible moves. The skill’s gate list also includes compliance, public policy, rights, employment, education, health, money, and physical safety. of1.ai states the irreversible case in one line: anything you cannot undo waits at a human gate.

OpenAI is treating this model as Critical in cybersecurity and High for biological and chemical capability. On this reading, a named person stays on tasks in those two areas. The ratings are not a measurement of any one network.

One invoice, two steps

Stay with the example OpenAI gives. Do not add a result the company did not report.

Reading an invoice and drafting the vendor, the amount, and the date is usually a step you can throw away. Entering those details into accounting software can mean a draft row a person still has to post. It can also mean a posted entry, a changed payment detail, or a payment already sent. The second kind touches money, and it is hard to undo.

A kill test for “the model entered the invoice correctly” does not need a new benchmark. Open the invoice and the accounting record. If the vendor, the amount, or the date does not match, the claim is dead, whatever the model wrote in its log. The log and a summary of that log are one account when they come from the same run. The invoice is the other source. That is this reading of the method’s source-independence pass, not a new rule.

The model can prepare the row. A named person posts anything that moves money.

Five questions before the model finishes

Run this on one real task this week. It holds that task up to the gate list. It is not a new part of the method.

1. Does the task touch money, health, a legal matter, employment, safety, rights, or reputation? If it does, a named person signs before the change. So does a task in biology or chemistry. OpenAI rates the model High for biological and chemical capability.
2. Can you undo the result this week without asking anyone else? If you cannot, it waits at a human gate.
3. What result would prove the output wrong, and which document already in hand would show it? If you cannot name the document, the task is not ready to hand off.
4. Is the only record of what happened the model’s own log? If it is, you have one claim, not a check.
5. Could the task grant access, change a credential, or follow instructions that arrived inside a file or a page? OpenAI is treating this model as Critical in cybersecurity. That rating is the company’s, not a test of your systems. Keep those steps with a person.

If the answer to 1, 2, or 5 is yes, the model may draft. It does not finish.

What to do this week

• Pick one computer-use task you were about to assign. Write the reversible draft on one line and the step you cannot undo on the next.
• Put a person’s name on the second line.
• Write the kill test as one sentence that names a document you already have.
• Read 16.7% and 4.3% as company-reported results on OpenAI’s error-inducing prompts, not as the error rate on your task.

If you want Ask, Map, Move run live on a decision your team already faces, that is a consult: $400 an hour, in 5-hour blocks.