OpenAI launches GPT-6.1 Terra for cheaper visual and coding tasks

A budget-friendly AI release

OpenAI has officially introduced GPT-6.1 Terra, a new addition to its GPT-6.1 model family. This specific model is designed to balance high performance with lower operational costs for developers and businesses. It serves as a more affordable alternative to the company's more expensive high-end models.

The company is positioning Terra as a practical tool for handling repetitive, high-volume tasks. By offering a lower price point, OpenAI aims to make it easier for organizations to integrate AI into daily workflows without excessive spending. It is intended to fill the gap between basic models and the most powerful, costly options available.

Users can access the new model through ChatGPT Work and Codex. It is also available for integration into external applications via the OpenAI API. The release marks a shift toward making specialized capabilities, like visual understanding, more accessible to a wider range of users.

Improved visual and software skills

One of the primary strengths of GPT-6.1 Terra is its ability to process visual information and interact with software interfaces. This includes capabilities such as interpreting technical drawings, completing online forms, and editing videos. These features allow the model to go beyond simple text generation to perform more complex digital actions.

The model also shows significant improvements in software engineering and factual accuracy. In tests involving command-line tools, it achieved a 50.3 per cent success rate on Terminal-Bench 4.0. This is more than double the 21.5 per cent score previously recorded by the GPT-5.6 Terra model.

OpenAI also tested the model's ability to interact with desktop applications to complete multi-step tasks. On the OSWorld 2.1 benchmark, Terra scored best at the intermediate reasoning level. These results suggest the model is highly effective at navigating computer environments to get work done.

Significant cost reductions

The pricing for GPT-6.1 Terra is a major highlight of the announcement. The standard API pricing is set at $1 per million input tokens and $5 per million output tokens. These rates are 50 per cent lower than the prices for both GPT-5.6 Terra and GPT-6.1 Sol models.

These savings are particularly relevant for companies that need to process massive amounts of data. For example, a business could use the model to automatically extract information from incoming invoices and input it into accounting software. This automation reduces the need for manual data entry while keeping costs manageable.

OpenAI notes that users of ChatGPT Work and Codex can accomplish more tasks within their existing usage limits when using Terra. By leveraging a model that is about one-third the cost of its predecessor in some benchmarks, users can stretch their budgets further. This makes it a strategic choice for high-frequency, lower-complexity operations.

Accuracy and security measures

OpenAI has also focused on making the model more reliable and safer for enterprise use. Internal testing showed that the model's error rate dropped significantly when reasoning effort was increased. At its peak effort level, the error rate fell to just 4.3 per cent, down from 53.6 per cent for the older version.

The model is also designed to better adhere to specific user constraints and query limits. It shows a better ability to express potential barriers or limitations when a user provides a prompt. This helps ensure the AI stays within the boundaries set by the human operator.

Security remains a priority, with Terra showing resistance to prompt-injection attacks. These are malicious instructions intended to mislead the AI from its intended purpose. The model maintains the same protection features as the GPT-6 Astra model and is currently in the Preparedness state for cybersecurity.

Where Terra fits in the lineup

While GPT-6.1 Terra is a major step forward for mid-range tasks, it is not intended to replace the most powerful models. OpenAI clarifies that for the most complex software engineering problems, the GPT-6.1 Sol and GPT-6 Astra models remain superior. Terra is specifically optimized for the