The GPT-6.1 Terra model is now accessible to developers via OpenRouter and within GitHub Copilot’s CLI and app. The model, which OpenAI designed for visual reasoning and “computer use,” is intended to automate business workflows by interpreting screen content and taking action across various software applications.

In agentic performance tests, the model solved 50.3% of tasks on the Terminal-Bench 4.0 coding benchmark, compared to 21.5% for the GPT-5.6 Terra version. Its hallucination rate on the AA-Omniscience benchmark fell to 38% from 88%, though GitHub Copilot’s early tests found it took more time and tokens to render complex scenes than the previous version. Under its preparedness framework, OpenAI classified the model as “critical” in cybersecurity and “high” for biological and chemical capabilities.

The model is available through the OpenAI API at $1 per million input tokens and $5 per million output tokens, with a 1.05M context window. OpenAI is also adjusting the default power presets in ChatGPT Work for some Plus and Pro users, substituting GPT-6.1 Sol at the “Light” setting with GPT-6.1 Terra at the “Medium” setting.