OpenAI introduced GPT-6.1 Terra on October 26, 2026, replacing GPT-5.6 Terra with a model designed for workflow automation, software interaction, visual analysis and text generation. The company positions Terra as a less expensive alternative to GPT-6.1 Sol and GPT-6 Astra, with comparable results on some tasks at a significantly lower cost.

The model is available to Plus, Pro, Enterprise and Edu subscribers through ChatGPT Work and Codex. Developers can access it through the gpt-6.1-terra API.

Features and test results

GPT-6.1 Terra can take actions in applications, interpret technical drawings, work with documents, complete forms, process invoices and interact with software through graphical interfaces.

In OpenAI’s Agents’ Last Exam benchmark, Terra at an average reasoning level outperformed GPT-5.6 Terra by 15 percentage points while costing about three times less. OpenAI said the result was comparable to GPT-6 Astra at one-tenth of the token cost. Terra also completed tasks in the benchmark more than three times faster than GPT-6 Astra and GPT-6.1 Sol.

On MMMU-Pro, the model outperformed GPT-5.6 Terra across all reasoning modes and approached the results of Astra and Sol. Its successful completion rate on Terminal-Bench 4.0 reached 50.3%, compared with 21.5% for the previous model.

OpenAI’s internal testing found that the share of factual errors at the minimum reasoning level fell from 53.6% to 16.7%. At the maximum reasoning level, the error rate was 4.3%. The company published further results for the Agents’ Last Exam benchmark.

OpenAI also focused on resistance to malicious instructions. According to the company, Terra follows user-request constraints more closely and is less likely to take actions outside its assigned task.

Under OpenAI’s internal Preparedness Framework, however, the model reached the Critical level in the cybersecurity category. It uses protective mechanisms developed for GPT-6 Astra.

API access costs $1 per 1 million input tokens, $0.05 per 1 million cached tokens and $5 per 1 million output tokens. Those rates are 50% lower than for GPT-5.6 Terra and GPT-6.1 Sol.

Independent evaluations

Artificial Analysis compared Terra and Sol at a medium reasoning level. Terra scored 41 on the research platform’s composite Intelligence Index, against 48 for Sol.

Terra generated 91 tokens per second, compared with 54 for Sol. The estimated cost of completing one task in the platform’s tests was $0.10 for Terra and $0.21 for Sol.

Plotly also tested the model on data-analytics tasks. Terra completed 40 of 43 tasks, achieving 93% accuracy, at a total testing cost of $1.31. GPT-6.1 Sol completed all 43 tasks for $2.61.

Plotly said Terra was the fastest model among those tested that correctly completed at least 40 tasks. “None of the models we tested turned out to be both cheaper and more accurate,” Plotly’s assessment says.

The evaluation also identified limits on the most difficult tasks. Terra found three of five hidden patterns in the data, while Sol identified all five. The results indicate Terra’s advantage is primarily in large-scale automation and everyday workflows where speed and cost are priorities, while OpenAI continues to recommend GPT-6.1 Sol and GPT-6 Astra for the most complex programming and analysis tasks.

OpenAI had previously faced safety questions involving autonomous AI agents after an experimental system breached its isolated environment. The company is placing additional emphasis on limiting model actions and improving resistance to attempts to manipulate instructions.

Source: Incrypted