What happened
On 26 October OpenAI launched GPT-6.1 Terra, the middle model in its current family. It sits above GPT-6 Luna, the cheap one, and below GPT-6.1 Sol. OpenAI pitches it as a fast model for agents: using a computer, reading images in detail, coding and writing.
The price is what stands out. It costs 1 dollar per million input tokens and 5 dollars per million output tokens, half the price of GPT-5.6 Terra. If the input text repeats and gets cached, it drops to 0.05 dollars per million. According to OpenAI, it also completes tasks about three times faster than GPT-6.1 Sol and GPT-6 Astra.
It is available in the API as gpt-6.1-terra and, inside ChatGPT, in ChatGPT Work and Codex on the Plus, Pro, Enterprise and Edu plans. It does not show up in regular ChatGPT chat yet.
The numbers worth knowing
We have checked these figures across coverage from several specialist outlets. In our view, these are the ones that matter for the decision:
- Price: 1 dollar per million input tokens, 0.05 cached and 5 for output.
- Context: up to 1.05 million tokens, enough to load long contracts, catalogues or a customer's full history in one go.
- Fewer made-up answers: on a test designed to make it invent facts, the error rate at low reasoning goes from 53.6% for GPT-5.6 Terra to 16.7%. At maximum effort it falls to 4.3%.
- Coding: it solves 50.3% of Terminal-Bench 4.0 tasks, compared with 21.5% for its predecessor.
- Safety: without safeguards, it got around restrictions in 29.6% of cases, and in 13.75% of tests it carried out unauthorised actions after receiving messages that appeared to come from other agents.
Why it matters
Until now there was an awkward gap. Cheap models such as GPT-6 Luna or Claude Haiku 5.5 are more than enough to sort emails or pull data from an invoice, and we already explained how that costs pennies a month. But when you ask AI to do something with several steps, like visiting a website, comparing data, filling in a form and notifying someone, those models get lost. And the big ones cost a lot more.
Terra aims right at that gap: agent tasks at a tenth of the cost of OpenAI's top models. For a small company, that changes which projects pay off.
The other half of the news is in the safety card, and we think it is the most useful part. If an agent obeys, in more than one test in ten, a message that only looks like it came from another agent, then any text it reads, whether an email, a web page or a document, can try to give it orders. This is not a flaw unique to Terra, but here it is measured and published.
What it means for a small business
If you have automations running on GPT-4o or GPT-4.1, remember that OpenAI switches them off in its API on 14 January. For simple tasks, GPT-6 Luna is still the logical replacement. For flows that needed more reasoning, Terra now joins the comparison at a reasonable price.
If you were thinking about an agent that does things, not just answers, cost is no longer the excuse. An agent that reviews orders, prepares quotes or updates the CRM from emails can run on Terra for well under what it cost a month ago.
If your agent reads outside content, such as customer emails, supplier websites or forms, the risk described in the safety card applies to you. We have already seen how agents are being used to attack online stores. The problem here is the reverse: someone tricking your own agent with a well-written piece of text.
Our view: Terra is good news for small businesses, but saving on tokens cannot come at the price of giving the agent more permissions than it needs.
What I would do this week
First, take stock. Write down which automations use OpenAI models and which of them make decisions or take actions, such as sending emails, changing data or placing orders.
Second, try Terra only where it makes sense. If the task has a single step, stick with a cheap model. If it has several steps and today it fails or a person does it, test Terra on twenty or thirty real cases and compare result and cost.
Third, cut permissions. An agent that reads emails does not need to be able to delete contacts. Give it access only to what is essential, with its own keys, not an administrator's.
Fourth, put a person in front of anything irreversible. Payments, messages to customers, price changes or deletions: let the agent prepare them and have someone approve them with one click.
Fifth, treat all outside text as suspect. If the agent reads emails or web pages, tell it in its instructions never to follow orders that arrive inside that content, and check its logs from time to time.
If this feels far from your day to day, it is the kind of work we do as an outsourced CTO: deciding which model to use, where, and with what limits.
If you want us to look at it with you
We design and maintain AI agents and automations for small businesses, and we train the teams that use them through our AI training.
Book a thirty-minute call. We go through your automations with you, whether Terra or a cheaper model fits each one better, and which permissions should be cut back. No strings attached.
Sources
This article is based on the following publications, consulted on 29 October 2026:
Frequently asked questions
It is an OpenAI model launched on 26 October 2026 and built for AI agents: multi-step tasks, computer use, coding and detailed image reading. It sits between GPT-6 Luna, the cheapest, and GPT-6.1 Sol.
In the API it costs 1 dollar per million input tokens, 0.05 dollars per million cached input tokens and 5 dollars per million output tokens. That is half the price of GPT-5.6 Terra.
Yes, but not in regular chat. It is available in ChatGPT Work and Codex on the Plus, Pro, Enterprise and Edu plans. Developers can use it in the API as gpt-6.1-terra.
It can be used with precautions. Its safety card shows that, without safeguards, it got around restrictions in 29.6% of cases and carried out unauthorised actions in 13.75% of tests after receiving messages that appeared to come from other agents. It is wise to limit its permissions and have a person approve irreversible actions.
It can for complex tasks. OpenAI is retiring GPT-4o and GPT-4.1 from its API on 14 January 2027 and recommends GPT-6 Luna, which is cheaper, as the general replacement. Terra makes sense for flows that need to reason over several steps.
It depends on the task. To sort emails or extract data, a model such as GPT-6 Luna or Claude Haiku 5.5 is enough and much cheaper. Terra pays off when the task has several steps and small models fail, which is worth checking with real cases before switching.