OpenAI on October 26 added a new middle tier to its GPT-6 family. GPT-6.1 Terra replaces GPT-5.6 Terra as the company's fast, lower-cost option, and OpenAI is selling it mainly on its ability to operate software: reading screens, filling in forms, interpreting technical drawings and working through the routine tasks of an office day. Plus, Pro, Enterprise and Edu subscribers can use it from launch in ChatGPT Work and Codex, and developers can call it in the API under the identifier gpt-6.1-terra.
Price is the headline for builders. On the Standard tier, a million tokens costs $1 going in and $5 coming out, with cached input billed at five cents per million. OpenAI's pricing table lists GPT-5.6 Terra at $2 and $12, and GPT-6.1 Sol at $2 and $10, so input costs are halved and output costs fall by more than half. Requests longer than 272,000 input tokens are billed at twice the input rate and 1.5 times the output rate, while Batch and Flex processing cost 50% less than Standard.
The developer documentation describes a model that takes text and images and returns text only, with no audio or video support. It has a context window of 1,050,000 tokens, accepts at most 922,000 input tokens, can produce up to 128,000 output tokens and has a knowledge cutoff of June 1, 2026. Reasoning effort can be set from none to max, and tool calling requires the Responses API.
Most of the performance case rests on OpenAI's own measurements. On Agents' Last Exam, a set of professional computer tasks, the company says Terra at medium reasoning effort scores 15 percentage points above its predecessor for roughly a third of the cost and matches GPT-6 Astra at a tenth of the per-token price. At maximum effort, OpenAI calculates a cost per task about nine times lower than Anthropic's Claude Opus 5.5, and it says Terra completes its work in under a third of the time that GPT-6 Astra or GPT-6.1 Sol need. It also reports gains on MMMU-Pro for visual reasoning and on OSWorld 2.1 for desktop-app tasks.
Coding and accuracy are the other claimed improvements. OpenAI puts Terra at 50.3% on Terminal-Bench 4.0 at maximum effort, against 21.5% for GPT-5.6 Terra, though it still recommends GPT-6.1 Sol and GPT-6 Astra for the hardest programming work. On an internal set of prompts built to provoke mistakes, the reported error rate at low effort drops from 53.6% to 16.7%, and to 4.3% at maximum effort.
On safety, OpenAI says the model reaches the Critical threshold for cybersecurity under its Preparedness Framework and inherits the safeguards built for GPT-6 Astra. It also cites better prompt-injection resistance and a stronger tendency to stay within the scope of a task, with details in a system card.
Distribution began immediately beyond OpenAI's own products. Amazon Web Services announced general availability of GPT-6.1 Terra on Amazon Bedrock the same day; its notice describes support for up to 1 million input tokens, a looser figure than the 922,000-token input cap in OpenAI's documentation. The Verge covered the launch largely through OpenAI's own description of the model.
The practical significance is cost. If the benchmark results hold up outside OpenAI's tests, Terra would make agent workloads that click through applications or process documents at volume considerably cheaper to run. For now, every comparison with rival models comes from OpenAI, and no independent evaluation had been published by the time of writing.