OpenAI has released a new version of the model most people will actually end up using at work. GPT-6.1 Terra, which went live on Monday 26 October, is the middle model in the company's range: cleverer than the small Luna, cheaper and quicker than the heavyweight Sol and Astra. It replaces GPT-5.6 Terra, and the headline is the price. Developers pay half what they did before, while OpenAI claims the model does far more.
The pitch is "computer use": an AI that looks at a screen, works out what is on it, and then clicks, types and moves between apps much as a person would. Terra is aimed squarely at the dull but valuable jobs that fill an office day.
What Terra is for
OpenAI describes Terra as a fast, efficient model that combines detailed visual reasoning, precise computer use and clearer, more natural writing. Its examples are deliberately ordinary: drafting proposals, interpreting technical drawings, filling in forms and editing video. The company's own illustration of a good use is a stream of incoming invoices, with the model reading each one and keying the details into accounting software.
It accepts text and images, and Microsoft says it handles up to one million input tokens in a single request.
In ChatGPT, Terra is available from today to Plus, Pro, Enterprise and Edu subscribers in ChatGPT Work and in Codex, OpenAI's coding tool. ChatGPT Work is the agent mode that can work across apps and websites on your behalf. Free users, according to Decrypt, stay on the smaller GPT-5.6 Luna.
The numbers, and who produced them
The price first. Standard API rates are $1 per million input tokens and $5 per million output tokens, with cached input at five cents. The Decoder, a specialist AI news site, puts GPT-5.6 Terra at $2 and $12, and Anthropic's mid-range Claude Sonnet 5.5 at $2 and $10. OpenAI says Terra's rates are 50% below both GPT-5.6 Terra and GPT-6.1 Sol.
On capability, OpenAI leans on a benchmark called Agents' Last Exam, which tests professional tasks such as turning architectural drawings into a 3D model. At medium effort Terra scores 15 percentage points above its predecessor for about a third of the cost, and OpenAI says it matches GPT-6 Astra, its most capable model, at a tenth of the price per token. Against Anthropic's Claude Opus 5.5 at maximum effort, OpenAI says Terra uses over three times fewer tokens at a quarter of the price per token, which works out at roughly nine times lower cost per task. It also finishes tasks more than three times faster than Astra and Sol.
The factuality claim is arguably the most important. On an internal test built from prompts designed to trick models, the error rate at low reasoning effort fell from 53.6% to 16.7%, and at maximum effort it dropped by 18 points to 4.3%. For coding, Terra solved 50.3% of tasks on Terminal-Bench 4.0 at maximum effort, against 21.5% for GPT-5.6 Terra. OpenAI still points users to Sol and Astra for the hardest programming work.
Treat all of that with care. These are OpenAI's own results, the error test is an internal one, and The Next Web notes the comparisons with Claude have not been independently checked. There is some outside support. Artificial Analysis, an independent benchmarking firm, scores Terra at 48 on its Intelligence Index, six points above GPT-5.6 Terra, while costing less than a third as much per task ($0.40 against $1.40). On its coding index Terra scores 48, six points behind Sol and 17 ahead of Luna. On a private test of hallucination, Terra guessed at an answer it did not know 38% of the time, down from 88% for its predecessor. Better, but a model that still guesses a third of the time needs checking.
Safety: a mid-range model with top-tier risk
One detail is easy to miss. OpenAI says Terra meets the "Critical" threshold for cybersecurity under its Preparedness Framework, the system it uses to rate dangerous capabilities, and that it carries the same safeguards as GPT-6 Astra. In other words, even the everyday model now has cyber skills that OpenAI itself labels "Critical". The system card reports that defences against indirect prompt injection, where malicious instructions are buried in a web page or document the model reads, held 98.8% of the time, up from 96.2% for GPT-5.6 Terra.
The launch also follows a bumpy month. The Next Web reports that OpenAI cancelled GPT-6.1 Astra last month after it failed internal safety tests, and we have covered the company's decision to train its frontier models offline after one slipped onto the live web. OpenAI says Terra is more likely than its predecessor to stay within the task it is given and to be honest about its limits.
The UK angle
For British businesses, the interesting part is cost. Plenty of UK firms spend their days on exactly the jobs Terra is pitched at: invoices, forms, catalogue updates and routine paperwork. Halving the price of the model that does that work lowers the bar for a small firm to try it. Prices are quoted in US dollars, and we have not seen sterling pricing, so UK customers should expect exchange rates and VAT to matter when they budget.
Where the model runs also matters. Microsoft says Terra is generally available on Foundry, with Data Zone deployments for the US, EU and Asia-Pacific. Its price list shows the EU Data Zone costs more than the global rate: $1.20 per million input tokens and $6 per million output tokens, against $1 and $5. The page does not name a UK-specific zone, so a British company that wants its data kept in Europe would need to look at the EU option and its premium. Amazon has also made Terra available on Bedrock.
The model's safety profile has a UK dimension too. OpenAI is one of the two labs promising credits to Britain's new AI Biosecurity Lab, part of the £600m AI package the Chancellor will put into Wednesday's Budget, which also includes an extra £120m for the AI Security Institute. A model OpenAI rates "Critical" for cyber, yet prices low enough for almost any business to use, is the kind of thing that body is likely to be asked about.
What's next
Rivals may respond on price: on the figures above, Anthropic's Sonnet line is now twice as expensive per input token. Independent testers will also get to work on the claims that rest only on OpenAI's own numbers, particularly the cost-per-task comparison with Claude. For a UK business, the sensible approach is a small pilot with a human checking the output, not a handover of the books.