On October 26, OpenAI announced "GPT-6.1 Terra," a new AI model suitable for image understanding, PC operation, and text generation. It will be available via API, as well as for Plus, Pro, Enterprise, and Edu users of ChatGPT Work and Codex. The standard API pricing has been set at half the cost of "GPT-6.1 Sol."
Terra is a model designed for tasks such as interpreting the content of drawings and product photos, or operating application screens to perform work. Its key feature is the ability to process tasks that combine image understanding and operation—for example, reading information from invoices and inputting it into accounting software—at a lower cost than higher-end models.
According to OpenAI, in the "Agents' Last Exam" evaluation, which includes practical challenges such as creating 3D models from blueprints, GPT-6.1 Terra achieved a score 15 points higher than the previous "GPT-5.6 Terra" when using medium inference strength, with the cost per task being approximately one-third. In the "OSWorld 2.1" PC operation evaluation, it also surpassed the previous Terra's highest score with medium inference.
Its code-writing capabilities have also improved. In "Terminal-Bench 4.0," which involves solving problems through command-line operations, it achieved a 50.3% success rate with the maximum inference setting, surpassing the previous Terra's published value of 21.5%. However, OpenAI explains that GPT-6.1 Sol and "GPT-6 Astra" are more suitable for the most complex coding tasks.
Regarding factual inaccuracies, an internal evaluation using prompts designed to induce errors showed that the error rate with a weak inference setting decreased from the previous Terra's 53.6% to 16.7%. With the maximum inference setting, the error rate was 4.3%.
Made with AI
