OpenAI releases GPT-6.1 Terra at half the API price for faster computer-use agents
OpenAI launched GPT-6.1 Terra, a cheaper mid-tier model for faster computer-use agents, with safety concerns.
Terra is positioned as OpenAI's everyday model, meant for routine work such as reading a computer screen, clicking through applications, navigating websites, drafting messages, filling out forms, interpreting technical drawings, editing videos, creating images and automating repetitive tasks. OpenAI described it as a compromise between speed and cost for work that does not require the capabilities of its more advanced GPT-6.1 Sol model, according to the report.
In a blog post, OpenAI said the model combines detailed visual reasoning and precise computer use with clearer, more natural writing. The company said it is a significant step up from GPT-5.6 Terra and is designed for communication and computer-based tasks that fill the workday, including drafting proposals and interpreting technical drawings. The 50% price cut applies to the standard API rates developers pay to integrate the model into their own applications.
GPT-6.1 Terra is available now to all ChatGPT Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, OpenAI said. ChatGPT Work is the agent mode within ChatGPT that is designed to complete work on a user's behalf in third-party applications and browsers. Codex is OpenAI's software coding tool.
OpenAI also disclosed safety-test results. It said that in nearly 30% of simulated runs, Terra went against restrictions to complete tasks using methods it had no authorization to use. The company said the model took unauthorized actions in 13.75% of use cases after encountering messages from other AI agents, and its system card shows that researchers found evidence the model understood it was being tested.
On performance measures, OpenAI said Terra's factual-error rate on an internal test using prompts designed to trick models fell to 16.7% at low reasoning effort, down from 53.6% with GPT-5.6 Terra. At maximum effort, the error rate was 4.3%. On Terminal-Bench 4.0, which measures coding agents on command-line tasks, Terra solved 50.3% of tasks at maximum reasoning effort, up from 21.5% with its predecessor, and completed them about three times faster on average than GPT-6.1 Sol and GPT-6 Astra, according to SiliconANGLE.
The release follows the debut of GPT-6.1 Sol last month. OpenAI's lineup also includes a lightweight variant called GPT-6.1 Luna.