- GPT-6.1 Terra focuses on visual reasoning and desktop application interaction.
- The new model significantly improves performance and reduces costs in benchmark tests.
- Terra's error rate is significantly lower than its predecessor, GPT-5.6 Terra.
- OpenAI reports enhanced resistance to prompt injection.
OpenAI has launched GPT-6.1 Terra, a new model designed to handle complex computing tasks at a lower cost. This model focuses on tasks that require visual reasoning and interaction with desktop applications. The release also significantly improves factual accuracy and programming performance, surpassing GPT-5.6 Terra. Terra targets practical workloads that enterprises wish to automate, avoiding high costs for each task, including processing invoices and interpreting technical drawings. OpenAI states that the model can handle demanding professional tasks at a fraction of the cost of its more expensive alternatives.
Performance Improvements and Cost Reductions
In terms of rapid performance, GPT-6.1 Terra has shown significant improvements on the Agents’ Last Exam benchmark, which covers complex professional tasks. Challenges include converting architectural blueprints into 3D models and recording individual video shots for repair teams. With moderate reasoning effort, Terra scored 15 percentage points higher than GPT-5.6 Terra. OpenAI achieved this improvement at about one-third of the cost. Terra also matched the performance of GPT-6 Astra in benchmark tests, but at approximately one-tenth the cost per token.
GPT-6.1 Terra is capable of using computers quickly and effectively. The model understands content on the screen, can operate between applications, naturally compose text, and has higher efficiency to meet demands throughout the workday. OpenAI states that Terra completes tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol. At maximum reasoning effort, its task costs are also about nine times lower than Claude Opus 5.5. These results extend to desktop automation.
OSWorld 2.1 evaluates the model's ability to operate desktop applications through multi-step workflows. Terra's performance with moderate reasoning effort surpassed the best scores of GPT-5.6 Terra, while significantly reducing the cost per task.
Enhanced Visual Reasoning Capabilities
Terra demonstrated strong visual reasoning performance in the MMMU-Pro test. In each test's reasoning setup, its scores exceeded those of GPT-5.6 Terra, while still being lower in cost and close to the scores of GPT-6 Astra and GPT-6.1 Sol. OpenAI has also worked to reduce factual errors that could undermine automated workflows. Its internal benchmark tests use prompts designed to expose weaknesses in the model's responses. At low reasoning effort, Terra recorded an error rate of 16.7%.
In contrast, GPT-5.6 Terra scored 53.6% in the same evaluation, representing a reduction of 36.9 percentage points. This improvement continued at maximum reasoning effort, with Terra's error rate dropping to 4.3%, an 18 percentage point decrease from its predecessor, bringing its performance closer to that of GPT-6 Astra and GPT-6.1 Sol.
Programming performance has also seen significant enhancements. Terminal-Bench 4.0 measures the ability of programming agents to solve command line issues. One example involves fixing web layouts that shift during loading. At maximum reasoning effort, Terra solved 50.3% of tasks in the benchmark, while GPT-5.6 Terra scored 21.5%, giving the new model a success rate more than double that of its predecessor. OpenAI still recommends using GPT-6.1 Sol and GPT-6 Astra for the most complex programming challenges.
GPT-6.1 Terra became available on October 27 for Plus, Pro, Enterprise, and Edu users through ChatGPT Work and Codex. Developers can also access the model identifier gpt-6.1-terra via the OpenAI API. Standard API pricing starts at US$1 per million input tokens (approximately HK$8). Cached input tokens are charged at US$0.05 per million (approximately HK$0), while output tokens are charged at US$5 per million (approximately HK$39).
OpenAI states that these prices are 50% lower than those for GPT-5.6 Terra and GPT-6.1 Sol. Terra retains the safety measures developed for GPT-6 Astra, with OpenAI reporting enhanced resistance to prompt injection and better disclosure of its limitations. Its readiness assessments have also placed the model at critical thresholds for cybersecurity. The lower pricing may make complex computational automation more accessible to developers and enterprises. However, GPT-6.1 Sol and GPT-6 Astra remain more powerful options for the most demanding programming tasks.
Item Specification Error Rate (Low Reasoning Effort) 16.7% Error Rate (Maximum Reasoning Effort) 4.3% Task Resolution Rate (Maximum Reasoning Effort) 50.3%
How GPT-6.1 Terra Changes Automation Workflows
The launch of GPT-6.1 Terra marks a significant advancement in the field of automated computing, particularly in visual reasoning and desktop application interaction. By reducing operating costs and improving accuracy, this new model enables enterprises to automate complex tasks more effectively, thereby enhancing productivity. With lower error rates, developers and businesses can rely on this technology for critical tasks with greater confidence. Additionally, enhanced safety measures provide users with greater assurance, making GPT-6.1 Terra an important choice for future automation solutions.

