OpenAI introduced GPT-6.1 Terra on its website on 26 October, making the model available through its API and in ChatGPT Work and Codex. OpenAI is pitching Terra for work that combines visual understanding with actions on a computer, at standard API prices it says are 50 per cent lower than those of GPT-5.6 Terra and GPT-6.1 Sol. For developers, the release offers a lower-priced option for building systems that work across computer-based tasks.
Key points
- At standard API rates, a million output tokens cost $5, a million input tokens $1 and a million cached input tokens $0.05.
- OpenAI reports a 15-percentage-point gain over GPT-5.6 Terra on Agents’ Last Exam v1 at medium reasoning effort, at about one-third the cost.
- On Terminal-Bench 4.0, OpenAI says Terra solved 50.3 per cent of tasks at maximum reasoning effort.
- OpenAI says the model meets the Critical threshold for cybersecurity under its Preparedness Framework.
Terra’s $1 input price
Developers can call the model through the OpenAI API as gpt-6.1-terra. OpenAI sets its standard API rate at $5 for a million output tokens, while a million input tokens costs $1 and a million cached input tokens costs $0.05. Plus, Pro, Enterprise and Edu users can also access Terra in ChatGPT Work and Codex; OpenAI says the same usage allowance there covers more tasks than it did with GPT-5.6 Terra, and substantially more than with GPT-6.1 Sol or GPT-6 Astra.
Terra sits between Luna and Sol in OpenAI’s model range, with Astra above them, 9to5Mac reported. The publication said OpenAI released updated Sol and Luna models in September without updating Terra. The new release therefore gives that middle tier a GPT-6.1 model alongside Sol, while leaving developers a choice between the tiers for different tasks.
The price of a completed task depends on more than the listed price per token. A model that uses fewer tokens to finish a job can cost less to run, just as a shorter journey can cost less even when the fare per mile is unchanged. OpenAI’s benchmark comparisons report both task scores and costs at specified reasoning-effort settings, which matters when judging its case for using Terra on repeated computer work.
Entering invoice details into accounting software could be a candidate for that kind of work: OpenAI names reading incoming invoices and entering their details as an example of the high-volume activity it wants developers to automate. If the model can handle the required reading and entry reliably, a lower cost per task would make repeated use less expensive. That outcome would still depend on how it performs on the invoices and software involved.
Agents’ Last Exam tests computer actions
OpenAI measured computer use on Agents’ Last Exam v1, whose professional tasks include making a 3D model from architectural drawings and producing a time-coded log of an animated film. At medium reasoning effort, the company reports that GPT-6.1 Terra scored 15 percentage points above GPT-5.6 Terra at about one-third the cost. OpenAI also says Terra performed on par with GPT-6 Astra at one-tenth its price per token in that test.
Those tasks put visual understanding and computer use together: the model must work from what it sees in a drawing or sequence of images and produce the requested result, rather than answer a question about the images alone. OpenAI says OSWorld 2.1 separately tests multistep work in desktop applications. At medium reasoning effort on that benchmark, the company reports that Terra exceeded GPT-5.6 Terra’s best score at a lower cost per task.
At maximum reasoning effort on Agents’ Last Exam, OpenAI says Terra used tokens more than three times as efficiently as Claude Opus 5.5 and cost one-fourth as much per token, yielding a roughly 9x lower cost per task. The company also reports that Terra finished tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol. Those comparisons concern the benchmark tasks and the effort settings OpenAI used, rather than a single fixed cost for every computer action.
OpenAI also tested visual reasoning with MMMU-Pro. It reports that Terra beat GPT-5.6 Terra at every reasoning setting and came close to the scores of GPT-6 Astra and GPT-6.1 Sol at lower cost. That measurement covers visual reasoning, while Agents’ Last Exam and OSWorld 2.1 include tasks requiring the model to work towards a result on a computer.
Terminal-Bench 4.0 and Terra’s safety tests
For coding, OpenAI tested agents on Terminal-Bench 4.0, which includes problems tackled through a command line, such as repairing a webpage whose layout moves as it loads. At maximum reasoning effort, the company reports that Terra solved 50.3 per cent of tasks, against GPT-5.6 Terra’s published score of 21.5 per cent. OpenAI still recommends GPT-6.1 Sol and GPT-6 Astra for the most complex coding tasks. AI Affairs previously reported OpenAI’s GPT-6 results on COBOL and other coding tests.
On OpenAI’s internal set of prompts designed to elicit errors, the company reports a 16.7 per cent error rate for Terra at low reasoning effort, compared with 53.6 per cent for GPT-5.6 Terra. At maximum effort, OpenAI says Terra’s error rate was 4.3 per cent. These are results on the company’s error-inducing prompts, with reasoning effort stated for each comparison.
OpenAI says GPT-6.1 Terra meets the Critical threshold in cybersecurity under its Preparedness Framework and retains protections developed for GPT-6 Astra. In its own prompt-injection tests — attempts to redirect the model with malicious instructions embedded in material it encounters — OpenAI reports that Terra resisted instructions intended to pull it away from the user’s task.
_-1.jpg){kind=link}