OpenAI Introduces GPT-6.1 Terra, a Faster Mid-Tier Model at Half the Price
Table of Contents
You might want to know
- What does GPT-6.1 Terra offer compared with its predecessor, and which users can access it?
- How should readers interpret OpenAI’s reported accuracy and benchmark results?
Main Topic
OpenAI released GPT-6.1 Terra on October 26 as an upgrade to the middle model in its lineup. The company describes Terra as an everyday model for routine work, including reading information displayed on a screen, interacting with applications, and drafting replies. Its release centers on two claims: developers can access the model through standard API pricing at half the price of GPT-5.6 Terra, and it can perform computer-based tasks more quickly and accurately.
Computer use refers to an AI system interpreting what appears on a screen and then clicking or typing to carry out a task. In practical terms, this can include moving between applications, navigating websites, and completing actions that would otherwise require a person to use a mouse and keyboard. These capabilities are distinct from simply answering a question in a chat: the model must interpret an interface and act within it.
The announced price reduction applies to standard API prices, which are the rates developers pay to connect a model to their own applications. It does not mean that every product or subscription has been discounted by the same amount. OpenAI characterized the launch as a significant step up over GPT-5.6 Terra while highlighting the 50% reduction in standard API prices. The company said the model is rolling out to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, and developers can access it through the API as gpt-6.1-terra.
ChatGPT Work is OpenAI’s agent mode within ChatGPT. It can work across applications, browse websites, and complete tasks on a user’s behalf. Codex is OpenAI’s coding tool. The source article says free users can use the lower-tier GPT-5.6 Luna, rather than GPT-6.1 Terra through the access routes described for paid and organizational users. Availability therefore depends on the product and account tier, as well as whether someone is using OpenAI’s products or integrating the model through the API.
OpenAI reported improvements in an internal test designed around prompts intended to trick models into making mistakes. At low reasoning effort—a setting that controls how much time the model spends thinking—the rate of factual errors fell from 53.6% for GPT-5.6 Terra to 16.7% for GPT-6.1 Terra. At maximum reasoning effort, the error rate fell by 18 points to 4.3%. These results suggest that the model’s factual reliability improved on the particular test OpenAI used, with the reported error rate lower at both settings.
These accuracy figures come from an internal OpenAI test, not an independently administered evaluation. That distinction matters when comparing models. The results describe how GPT-6.1 Terra performed on a test assembled by its maker, and readers should not assume that the same error rates will apply across every subject, task, or real-world setting. Test design, prompt selection, and the definition of a factual error can all affect a measured result.
OpenAI also reported results on Terminal-Bench 4.0, a benchmark that evaluates coding agents on tasks performed through a computer’s command-line terminal. It scores the share of tasks completed successfully. At maximum reasoning effort, GPT-6.1 Terra solved 50.3% of the tasks, compared with 21.5% for GPT-5.6 Terra. This is a sizable reported difference on that benchmark, though it represents performance on terminal-based tasks rather than a general measure of all coding ability.
The company further said GPT-6.1 Terra finishes tasks more than three times faster on average than GPT-6 Astra and GPT-6.1 Sol. The source does not provide additional details about the task mix or measurement method behind this speed comparison. As a result, it is useful to treat the statement as an OpenAI-reported comparison, rather than as a universal guarantee that Terra will be faster for every workload.
The model’s position in OpenAI’s lineup helps explain its intended role. Terra is presented as a practical option for regular office and computer-use tasks, while OpenAI continues to direct users toward GPT-6.1 Sol or GPT-6 Astra for in-depth research or complex coding. The distinction suggests that the company is positioning models around different levels and types of work: everyday actions for Terra, and more demanding research or programming for the other models.
The release also follows earlier milestones in the Terra line. OpenAI introduced Terra on June 26 as part of the GPT-5.6 family in a limited preview requested by the Donald Trump administration. The company opened it to the public on July 9, presenting it as competitive with GPT-5.5 at about half the cost. On September, OpenAI launched GPT-6 Sol and Luna minutes after Anthropic released Claude Opus 5.5, without introducing a new Terra at that time. GPT-6.1 Terra is therefore a later update to the middle-model offering, rather than the first appearance of the Terra name.
For developers, the combination of a lower standard API price and stronger results on OpenAI’s reported tests may make the model worth evaluating for tasks involving user interfaces, routine workflows, and terminal-based coding. The practical value will depend on how a particular application balances cost, response time, accuracy, and the consequences of mistakes. For businesses and individual users, the model’s availability in ChatGPT Work and Codex may make it possible to try its computer-use capabilities without building a separate integration.
However, benchmark results alone do not establish whether a model is suitable for a specific organization or task. Teams considering computer-use systems generally need to assess how a model handles unfamiliar interfaces, ambiguous instructions, sensitive data, and actions that are difficult to reverse. They may also want to confirm whether the model’s performance remains consistent when tasks differ from the examples used during testing. The reported results provide a starting point for evaluation, not a substitute for testing the model in the intended environment.
Key Insights Table
| Aspect | Description |
|---|---|
| Release and purpose | OpenAI released GPT-6.1 Terra on October 26 as a mid-tier model for routine office work and computer-use tasks. |
| API pricing | Standard API prices are 50% lower than those of GPT-5.6 Terra. |
| Availability | The model is rolling out to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex. Developers can use gpt-6.1-terra. Free users can use GPT-5.6 Luna. |
| Reported factual-error rates | In OpenAI’s internal test, errors at low reasoning effort fell from 53.6% to 16.7%; at maximum effort, the rate fell 18 points to 4.3%. |
| Terminal-Bench 4.0 | At maximum effort, GPT-6.1 Terra solved 50.3% of tasks, versus 21.5% for GPT-5.6 Terra. |
| Speed and model selection | OpenAI says Terra finishes tasks more than three times faster on average than GPT-6 Astra and GPT-6.1 Sol, while recommending those models for in-depth research or complex coding. |
| Evidence caveat | The factual-error test and reported comparisons are OpenAI’s own; the error test is internal. |
Afterwards...
GPT-6.1 Terra’s launch reflects a broader shift toward AI systems that do more than generate text. Models that can interpret screens, use applications, browse websites, and operate command-line tools may become useful components of everyday workflows. As these systems spread, progress will depend not only on raw task completion but also on whether their actions are reliable, understandable, and appropriate for the context.
Future evaluations could help users compare models more fairly by using independent tests, clearly documented task sets, and consistent definitions of success and error. For computer-use systems, useful research areas include how models recover from mistakes, recognize when they are uncertain, protect sensitive information, and request confirmation before taking consequential actions. These capabilities can matter as much as speed when an AI is permitted to interact with real applications or data.
Organizations adopting such tools may also need practical safeguards: limited permissions, review steps for high-impact actions, clear records of what an agent did, and ways for a person to interrupt or correct it. These measures do not eliminate every risk, but they can make automated workflows easier to supervise. The central question for the next generation of AI is not only what tasks a model can complete, but how reliably and transparently it completes them.
As models continue to develop, users and researchers will benefit from examining price, speed, accuracy, and safety together. GPT-6.1 Terra’s lower API pricing and improved results on OpenAI’s reported tests make it a notable update in the company’s lineup. Independent evaluation and real-world testing will help determine how well those advantages carry over to the varied tasks people and organizations actually need to perform.
Last edited at：2026/10/27