OpenAI has released a "cost-effective" AI model that offers improved performance relative to its cost. Developers can utilize it to automate large volumes of work at low costs, such as reading received invoices and entering detailed figures into accounting software.
On the 27th, OpenAI announced the release of 'GPT-6.1 Terra,' an AI model that enhances capabilities in everyday tasks and computer operations, including drafting proposals, interpreting technical drawings, form filling, and video editing. The standard API (Application Programming Interface) price is 50% lower than that of GPT-5.6 Terra and GPT-6.1 Sol, and the model is available through ChatGPT Work, Codex, and the OpenAI API.
In 'Agents' Last Exam,' which consists of professional tasks such as converting architectural drawings into 3D models or reviewing animations scene by scene to create timecode logs, GPT-6.1 Terra recorded a score 15 percentage points higher than GPT-5.6 Terra at the medium reasoning level. The cost was approximately one-third that of the predecessor. Compared to GPT-6 Astra, it demonstrated similar performance while offering a price per token that was one-tenth as high.
In MMMU-Pro, an evaluation of visual reasoning, GPT-6.1 Terra outperformed GPT-5.6 Terra across all reasoning settings. In OSWorld 2.1, which evaluates multi-step task capabilities using desktop apps, it surpassed the highest score of GPT-5.6 Terra at the medium reasoning level, while also lowering the cost per task.
Hallucinations and factual errors have also decreased. In an internal evaluation consisting of questions designed to induce errors, when the reasoning level was set low, GPT-6.1 Terra's error rate was 16.7%, lower than that of GPT-5.6 Terra (53.6%). When the reasoning level was set to maximum, the error rate was 4.3%, which is 18 percentage points lower than the previous model and comparable to GPT-6 Astra or GPT-6.1 Sol.
Coding performance has also improved. In Terminal-Bench 4.0, which evaluates the problem-solving capabilities of coding agents in a command-line environment, GPT-6.1 Terra solved 50.3% of tasks at the maximum reasoning level. This figure is more than double the publicly reported score of 21.5% for GPT-5.6 Terra. However, OpenAI explained that "GPT-6.1 Sol and GPT-6 Astra are better suited for complex coding tasks."
GPT-6.1 Terra is available to Plus, Pro, Enterprise, and Edu users of ChatGPT Work and Codex. Developers can access it as 'gpt-6.1-terra' via the OpenAI API. The standard API fee is $1 per 1 million tokens for input, $0.05 for cached input, and $5 for output.
For more details, please refer to the official OpenAI research blog.
in TECH
