OpenAI has released GPT-6.1 Terra, a fast model aimed at computer use and the document, form and coding work of a normal workday. Its standard API prices are 50 percent below those of GPT-5.6 Terra and GPT-6.1 Sol. In its announcement, OpenAI gives Terra a 50.3 percent score on Terminal-Bench 4.0 at maximum reasoning effort, more than double the 21.5 percent it published for GPT-5.6 Terra.
Terra costs $1 per million input tokens and $5 per million output tokens, with cached input at $0.05
The model is available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and through the OpenAI API as gpt-6.1-terra. OpenAI says the same plan allowance in ChatGPT Work and Codex covers more tasks than with GPT-5.6 Terra, and substantially more than with GPT-6.1 Sol or GPT-6 Astra.
At those rates output costs five times as much as fresh input, and cached input costs one twentieth as much as fresh input. A call with 100,000 fresh input tokens and 10,000 output tokens comes to $0.15; if the 100,000 input tokens come from the cache, it comes to $0.055. For agent loops that resend a long context, the cache rate matters more than the headline price.
The Terminal-Bench 4.0 score more than doubles and the error rate on OpenAI’s own prompts drops to 16.7 percent at low effort
OpenAI reports that Terra solves 50.3 percent of Terminal-Bench 4.0 tasks, which are command-line jobs such as fixing a webpage whose layout shifts as it loads, and adds that GPT-6.1 Sol and GPT-6 Astra remain the better choice for the most complex coding. On Agents’ Last Exam, a benchmark of professional computer-use tasks, Terra at medium effort scores 15 points above GPT-5.6 Terra at about a third of the cost, and matches GPT-6 Astra at one tenth of Astra’s price per token. It completes tasks more than three times faster than Astra and Sol.
On an internal benchmark of prompts built to induce errors, the error rate at low effort fell from 53.6 percent with GPT-5.6 Terra to 16.7 percent. At maximum effort the error rate fell by 18 points to 4.3 percent, close to the rates of Astra and Sol. On Terminal-Bench-Science 0.1, Terra surpasses Sol at lower cost and comes within five points of Claude Opus 5.5 at less than a fifth of the estimated cost. For MMMU-Pro and OSWorld 2.1 the announcement gives only the direction, with the scores in charts.
The nine-fold cost advantage over Opus 5.5 does not follow from the two figures given for it
At maximum effort OpenAI describes Terra as more than three times as token efficient as Claude Opus 5.5 at one quarter of its price per token, and puts the result at about 9x lower cost per task. Three times fewer tokens at a quarter of the price could multiply to 12x, and the announcement does not show how it arrived at 9x.
The error-rate benchmark is OpenAI’s own, and the announcement cites no independent test of Terra. It gives no context window or output limit. It does state that Terra meets the Critical threshold for cybersecurity under OpenAI’s Preparedness Framework, and that it keeps the protections developed for GPT-6 Astra.