OpenAI Launches GPT-6.1 Terra: Fast Computer-Operating Model at Half the Cost
OpenAI launches GPT-6.1 Terra to automate desktop software at 50% lower API pricing, meeting its Critical cybersecurity threshold alongside GPT-6 Astra.
OpenAI has officially launched GPT-6.1 Terra, a high-speed frontier AI model specialized in operating desktop software and executing agentic computer workflows. Released across ChatGPT Work, Codex, and the developer API at 50% lower token pricing than previous tiers, Terra becomes only the second OpenAI model to meet the 'Critical' cybersecurity threshold under its Preparedness Framework.
Desktop Computer Operation at Half the Token Cost
On October 27, 2026, OpenAI deployed GPT-6.1 Terra to all Plus, Pro, Enterprise, and Edu subscribers within ChatGPT Work and Codex, alongside immediate API access as gpt-6.1-terra. Purpose-built to autonomously navigate, click, and execute complex workflows inside native desktop applications—such as reading invoices, cross-referencing databases, and populating accounting software—Terra slashes API pricing by 50 percent compared to GPT-5.6 Terra and GPT-6.1 Sol to $1.00 per million input tokens, $0.05 per million cached inputs, and $5.00 per million output tokens.
Meeting the 'Critical' Cybersecurity Risk Threshold
Significantly, OpenAI designated GPT-6.1 Terra as meeting the 'Critical' cybersecurity risk threshold under its April 2025 Preparedness Framework, joining flagship GPT-6 Astra. The Critical rating denotes an autonomous capability to discover, model, and execute functional zero-day software exploits against hardened production environments without human intervention. To mitigate offensive misuse, OpenAI integrated specialized Astra-class guardrails, hardened prompt injection defenses, and architectural constraints designed to prevent unauthorized lateral network traversal.
Benchmark Leaps Across Real-World Agent Workflows
In performance evaluations on Agents' Last Exam, Terra scored 15 percentage points above GPT-5.6 Terra at one-third the execution cost, matching Astras capability at one-tenth the per-token price. In Terminal-Bench 4.0 command-line testing, Terra solved 50.3 percent of tasks (up from 21.5 percent in GPT-5.6 Terra), completing complex multi-step objectives more than three times faster than GPT-6.1 Sol while achieving nine-fold lower task costs compared to Claude Opus 5.5.