GPT-6.1 Terra is OpenAI’s new midrange model for visual tasks, writing, and automated software operation. It targets people and organizations that want to process large numbers of repetitive computer tasks without paying for a top-tier model every time. Its release also arrives amid a sharper price contest between OpenAI, Anthropic, and other providers.
Why Terra matters
OpenAI describes GPT-6.1 Terra as a model for fast visual reasoning, precise computer use, and more natural writing. A computer-use agent is an Artificial Intelligence (AI) system that can interpret user interfaces and carry out steps inside software. Rather than merely producing an answer, such an agent can fill in a form, move information into an application, or extract details from a technical drawing.
The official OpenAI announcement accordingly emphasizes fast and capable computer use. More detailed coverage lists drafting proposals, editing videos, and matching product photographs with catalog entries among the intended applications. These are not obscure demonstrations: they are recurring tasks in administration, sales, media production, and technical planning.
Terra accepts text and images but produces text as its output. It has a context window of 1.05 million tokens and can generate as many as 128,000 output tokens. Tokens are small units of text that a model uses to process input and output; a large context window can accommodate lengthy documents or extended work sessions.
The reports say the model is available through an Application Programming Interface (API), as well as in ChatGPT Work and Codex. An API is the connection other applications use to call an AI model and pay according to consumption. GPT-6.1 Terra is hosted exclusively by OpenAI and uses closed model weights, so you cannot operate it on your own infrastructure.
What becomes cheaper
Standard pricing is $1 per million input tokens and $5 per million output tokens. GPT-5.6 Terra cost $2 and $12 respectively, according to the report covering its price and performance figures. Cached input, which does not need to be processed from scratch each time it is reused, costs $0.05 per million tokens.
This makes Terra substantially cheaper than GPT-6.1 Sol and roughly half the price of Anthropic’s Claude Sonnet 5.5, whose listed rates are $2 for input and $10 for output. At the same time, Simon Willison notes that Terra costs about ten times as much as GPT-6 Luna. The choice is therefore not automatically simple: Terra sits between a very inexpensive model for basic jobs and more costly models for demanding work.
Computer-use agents benefit particularly from lower unit costs. An agent may analyze numerous screen views, plan intermediate actions, and generate repeated responses while completing a single assignment. If you process hundreds of forms or compare a large image collection with a product catalog, the relevant figure is not the cost of one request but the total cost of each completed task.
Published rates are also only one part of the current price competition. Reporting on individual OpenAI discounts describes a reduction of more than 30 percent for Replit and cost cuts of 80 to 90 percent for DataCamp. A U.S. bank reportedly received access for 500 employees instead of 250 at no additional charge.
Those enterprise agreements are not generally available price plans. Small teams should not assume that a 90 percent discount will appear merely because they ask nicely. The cases instead show that published API rates may not be the final offer at high volumes and that providers are currently willing to negotiate around volume, contract duration, and switching models.
Where the performance is sufficient
On Agents’ Last Exam, a performance test involving professional computer tasks, Terra scored 57.6 percent at medium reasoning effort, according to OpenAI, at a cost of $0.29 per task. GPT-5.6 Terra achieved 42.6 percent at $0.84, while GPT-6 Astra also scored 57.6 percent but cost $4.82. A benchmark is a standardized performance test; it supports comparisons but never fully represents your own daily work.
At maximum reasoning effort, Terra’s result increased to 60 percent and $1.01 per task. Claude Opus 5.5 achieved 63.2 percent in the reported comparison but cost $8.82. OpenAI also says Terra completes tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol. These figures come from provider claims and have not been independently confirmed for your particular use case.
The gap to more expensive models is narrow on visual tasks. Terra scored 85.5 percent on MMMU-Pro Vision at maximum reasoning effort, compared with 86.3 percent for GPT-6.1 Sol and 86.7 percent for GPT-6 Astra. The reported cost per test task was $0.012 for Terra and $0.065 for Astra.
This does not mean Terra is similarly competitive everywhere. It solved 50.3 percent of Terminal-Bench 4.0, which tests work in a text-based system environment. GPT-6.1 Sol reached 55.9 percent and Claude Sonnet 5.5 reached 70.6 percent. Simon Willison’s assessment of the model’s position likewise presents Terra as especially appealing for vision, computer use, and technical communication, but somewhat weaker on many other kinds of work.
OpenAI also reports lower error rates on an internal set of intentionally difficult prompts. At low reasoning effort, the rate reportedly fell from 53.6 to 16.7 percent, while maximum effort brought it down to 4.3 percent. Another private test covered by THE DECODER also points to fewer hallucinations, meaning plausible-sounding but false statements. Internal and private evaluations, however, do not replace testing with your own documents and workflows.
Pros and Cons of GPT-6.1 Terra
Pros:
- Lower cost – Input is half the predecessor’s price, while output falls from $12 to $5 per million tokens.
- Strong visual performance – Terra comes close to substantially more expensive models in some reported vision and computer-use tests.
- Greater speed – OpenAI says Terra handles computer tasks more than three times faster than Astra and Sol.
- Practical focus – Forms, proposals, technical drawings, and product images correspond to real office and professional work.
Cons:
- Not the strongest model – Top-tier alternatives remain ahead on terminal work and several computer-use evaluations.
- OpenAI dependency – Closed weights and exclusive OpenAI hosting rule out operation on your own infrastructure.
- Inconsistent discounts – Large enterprise deals reveal little about the price a small organization will actually receive.
- Limited evidence – Much of the performance data comes from OpenAI or private evaluations.
What this means for you
If you are getting started, first choose a tightly defined task that occurs frequently. Suitable examples include transferring details from standardized documents into a form or producing an initial proposal draft from supplied information. Compare Terra with a cheaper model, and measure not only writing quality but also cost, completion time, and errors per fully finished job.
Advanced users can assign different kinds of work to different models. An inexpensive option such as Luna might handle simple classification or short standard text, Terra can perform visual checks and operate software, and Sol or a competing model can be reserved for unusually difficult cases. This kind of model routing sends each task to an appropriate price and performance tier instead of processing everything with the same expensive model.
At higher consumption levels, you can use competition among providers in pricing discussions. To do so effectively, you need reliable numbers for monthly token consumption, successfully completed tasks, and the cost of switching providers. The reported discounts demonstrate room for negotiation, but they do not establish that every organization will receive comparable terms.
Users in Switzerland face the same unresolved questions as those elsewhere, and the sources do not list Swiss-specific prices or availability restrictions. Because Terra is hosted only by OpenAI, organizations working with confidential customer, employee, or education data need to determine whether the intended service fits their own privacy requirements. The supplied sources do not identify a hosting location in Switzerland.
GPT-6.1 Terra makes computer-use agents more economically attractive by delivering much of the reported top-tier performance on visual and common office tasks at a lower cost. It is still not a universal replacement for stronger or cheaper models, since the right choice depends on the job. The unresolved risk is that provider benchmarks, closed infrastructure, and individually negotiated discounts may look more favorable than your eventual bill and real-world error rate.
Sources
- OpenAI Launches GPT-6.1 Terra, Halving GPT-6.1 Sol’s API Price for Agent Work – MarkTechPost, 2026-10-27
- GPT-6.1 Terra – Simon Willison’s Weblog, 2026-10-26
- OpenAI bringt GPT-6.1 Terra: Günstiger und genauer als der Vorgänger – THE DECODER, 2026-10-26
- Introducing GPT-6.1 Terra – OpenAI, 2026-10-26
- KI-Preiskampf: OpenAI lockt Anthropic-Kunden mit Rabatten von bis zu 90 Prozent ab – THE DECODER, 2026-10-25
Image: Yan Krukau via Pexels