GPT-6.1 Terra is OpenAI’s mid-sized model, released on 26 October 2026, designed to work on a computer: it understands what it sees on the screen, switches between applications and fills in forms. It is available in Work mode and in Codex for Plus, Pro, Enterprise and Edu plans, as well as via the API, at half the price of its predecessor. For brands, this means that, increasingly, a bot will visit their website to compare offers on the customer’s behalf and submit an enquiry.
Below, we explain how GPT-6.1 Terra differs from other OpenAI models, where you can use it, what the tests have shown, and what this means for brand visibility in AI. This is a separate change from the one described in the text GPT-6 in ChatGPT: what it means for brand visibility: that one concerned the standard chat feature used by most customers, whereas Terra operates in Work mode, in Codex and in applications built using the API.
This is the mid-range model in the GPT-6 family, designed for speed and low cost. OpenAI describes it as a model that combines an accurate understanding of images with precise computer processing. Terra replaces GPT-5.6 Terra and sits between the most affordable GPT-6 Luna and the more powerful GPT-6.1 Sol and GPT-6 Astra. OpenAI released it on 26 October 2026.
One million input tokens in the API cost $1, whilst one million output tokens cost $5. This is half the price of GPT-5.6 Terra and GPT-6.1 Sol. The most expensive GPT-6 Astra costs $10 and $50, whilst the cheapest GPT-6 Luna costs $0.10 and $0.50.
In ChatGPT’s Work mode and in Codex, on the Plus, Pro, Enterprise and Edu plans, and for developers via the API under the name gpt-6.1-terra. OpenAI’s announcement makes no mention of the Free and Go plans or the standard chat tab. A user who asks ChatGPT for a business recommendation therefore continues to interact with the models described in the text about GPT-6. Terra is used where ChatGPT is required to perform a task, rather than simply providing a response.
This division of models between plans and modes is important when checking what AI chatbots select responses. The result depends on which model is used, so a test carried out on a paid account in ‘Work’ mode will not show the same results as what a customer would see in a free chat.
First and foremost, it operates a computer: it understands what is on the screen, switches between applications and carries out tasks within them. OpenAI gives examples from everyday work: filling in forms, video editing, reading technical drawings, writing quotations and automating repetitive tasks on a computer. According to OpenAI, it performs such tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol.
Wyniki testów podaje OpenAI, więc traktuj je jako deklarację producenta. W teście obsługi komputera OSWorld 2.1 Terra rozwiązała 74,2% zadań, a GPT-5.6 Terra 57,1%. Najlepszy w tym zestawieniu był Claude Sonnet 5.5 z wynikiem 83,9%. W teście zadań zawodowych Agents' Last Exam Terra osiągnęła 60,0% przy koszcie 1,01 dolara za zadanie, a GPT-6 Astra 59,3% przy koszcie 7,24 dolara. W programowaniu, w teście Terminal-Bench 4.0, wynik wzrósł z 21,5% do 50,3%.
According to OpenAI, yes. On an internal set of questions designed to induce errors in the model, the proportion of incorrect answers at a low level of reasoning fell from 53.6% in GPT-5.6 Terra to 16.7%. At the highest level of reasoning, it stood at 4.3%, which is only slightly higher than in GPT-6 Astra and GPT-6.1 Sol.
Our assessment: this is good news for companies about which models write untruths, but it’s no reason to stop checking. A model who makes things up less often is more likely to repeat what they find in their sources. If there are out-of-date prices or an old address online, they’re more likely to quote them. We explain what to do in such cases in our articles AI is providing false information about the company and What AI says about your brand’s weaknesses.
More and more often, your website will be visited by a bot rather than a human. A bot that reads the screen, clicks and fills in forms can open several pages on behalf of a customer, compare prices and send an enquiry. Our assessment: what matters here is the same as in AI chat responses – namely, specific facts stated clearly – plus something new: whether the bot will navigate through your website to the very end.
A website where the price is simply „negotiable” and the form contains unclear fields will lose out to a competitor where an agent can find the price, delivery date and terms and conditions in a matter of seconds. This is an extension of the principle described in the text Knows the brand, but doesn’t recommend it: the model selects a brand for which it has comparable data. We explain how this works in shops in the text An online shop in AI chatbots.
A lower price is also a factor. InfoWorld points out that cheaper tokens encourage large-scale automation, although analysts remind us that the cost of human verification of an agent’s work remains. Our assessment: the cheaper the model, the more companies will build their own assistants and shopping agents on top of it. These are further areas where the model determines which brand to display, alongside ChatGPT itself, AI mode on Google and AI Overviews.
Just like with AI chatbots, only more precise. The first step is access, then content that can be compared, and finally a path that the agent can follow without having to guess.
Access for the ChatGPT search bot is described in the OpenAI documentation, and we explain in this article how AI chatbots search for information online Measurement with web search. Straightforward facts are the foundation AEO, and structured data They help the models to correctly distinguish between the price and the product name. Among the technical extras, a file is also often included llms.txt, which we describe in the Dictionary.
The fact that this works can be seen in the example of a lime manufacturer. After adding standards, specifications and prices to its existing web pages, the brand’s ranking rose from 0 to 17 out of 18 queries in the July 2026 analysis. We describe the full story in the article A case study on visibility in AI, and the general guidelines in the guide How to get AI to recommend your company.
Not in a normal chat, because Terra didn’t end up there. If the brand’s results in ChatGPT changed towards the end of October, look elsewhere for the reasons, not in this model. AI chat responses change even without a new model, as we explain in the text Why does AI give a different answer every time?.
It’s worth noting down the date 26 October 2026 in your notes on the results, though. Some users on paid plans work in ‘Work’ mode, so they receive replies from a different model than in the chat. When comparing periods, keep the same frozen query set and look at the percentage of responses mentioning the brand, rather than its position on the list. We explain in the text how often to measure in order to distinguish between a change and noise How often should visibility be measured?, and what to do in the event of a sudden drop, as described in the text A decline in visibility in AI.
Aelo measures, diagnoses and draws up a plan. Module Visibility in AI Every day, it asks the AI chatbots the same set of customer questions and displays mention, citation and sentiment, so you can see straight away whether changing the model has shifted anything. Synapsis gathers facts about the brand that the models should know, and Tickets converts results into tasks: what to add to the page, where to correct the data, what gaps in the content fill in. You will prepare the content in Studio. We describe the whole process in the text An AI-powered brand visibility platform: how it works.
Agencies that explain the changes introduced with each new model to their clients can find out more on the website Aelo for agencies, and small businesses on the website Aelo for small businesses. We compare the cost of continuous monitoring in the text How much does visibility monitoring cost in AI?. You can take your first measurement on the Free account, without a card.
No. OpenAI makes it available in Work mode and in Codex as part of the Plus, Pro, Enterprise and Edu plans. The free version of ChatGPT uses the GPT-6 Luna model.
$1 per million input tokens and $5 per million output tokens. Tokens from the cache cost $0.05 per million.
Terra is half the price and, according to OpenAI, more than three times faster for computer-based tasks. For the most challenging programming tasks, OpenAI still recommends GPT-6.1 Sol or GPT-6 Astra.
Not in every respect. In the OSWorld 2.1 computer performance test, the Claude Sonnet 5.5 completed 83.9% tasks, whilst the Terra completed 74.2%. However, the Terra works out cheaper per task.
Yes, computer-assisted models are capable of filling in forms. That is why it is important for the fields to have clear labels, and for prices and terms and conditions to be written out in plain text.
