GPT-6.1 Terra meets OpenAI’s Critical cyber threshold, and it’s cheaper than GPT-5.6 Terra
OpenAI has released GPT-6.1 Terra, a model that can operate desktop software. It is now available to all Plus, Pro, Enterprise, and Edu users through ChatGPT Work and Codex. Developers can also use it through the API under the name gpt-6.1-terra. API prices are 50 percent lower than those for GPT-5.6 Terra and GPT-6.1 Sol: $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens. Tokens are small pieces of text that a model reads and writes.
OpenAI says Terra meets the Critical cybersecurity level in its Preparedness Framework. It also keeps the safety protections developed for GPT-6 Astra. OpenAI's own example of a task suited to Terra is reading large numbers of incoming invoices and entering their details into accounting software.
What Critical means
In the April 2025 version of OpenAI's Preparedness Framework, Critical is the higher of two cybersecurity risk levels. A model meets this level if, when given tools, it can find and create working attacks that exploit previously unknown software flaws, called zero-days, across all levels of severity in many well-protected, real-world critical systems without human help. It can also meet the level if it can plan and carry out a new attack from beginning to end against a well-protected target when given only a broad goal.
OpenAI's rules require the company to put safety measures in place during development, no matter how it plans to release a model. They also say that further development must stop until the company has defined safeguards that meet the Critical standard. On September 1, 2026, OpenAI classified Astra as its first model to reach this level.
Prompt injection
OpenAI says Terra performed well in its prompt injection tests. These tests used harmful instructions designed to make the model stop following the user's task and do something else. Compared with GPT-5.6 Terra, the new model is more likely to stay focused on the assigned task and to be honest about its limitations and anything that prevents it from completing the work.
What OpenAI measured
OpenAI tested Terra using Agents' Last Exam, a benchmark that measures performance on professional tasks. One example is turning architectural drawings into a 3D model. When using a medium level of reasoning effort, Terra scores 15 percentage points higher than GPT-5.6 Terra while costing about one-third as much. OpenAI says Terra performs as well as GPT-6 Astra while costing one-tenth as much per token.
At the highest reasoning effort, OpenAI says Terra uses tokens more than three times as efficiently as Claude Opus 5.5 and costs one-fourth as much per token. Together, these differences make the cost of completing a task roughly nine times lower. Terra also finishes tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol.
OpenAI also tested Terra with its own set of prompts designed to cause errors. At a low reasoning effort, Terra's error rate dropped from 53.6 percent with GPT-5.6 Terra to 16.7 percent, or about one error for every six prompts. At the highest reasoning effort, the error rate fell by 18 percentage points to 4.3 percent.
Hard coding tasks go elsewhere
On Terminal-Bench 4.0, a test of coding agents that covers command-line problems, Terra solves 50.3 percent of tasks at the highest reasoning effort. One example is fixing a webpage whose layout shifts while it loads. GPT-5.6 Terra has a published score of 21.5 percent on this benchmark. Terra still fails to solve 49.7 percent of tasks, and OpenAI says GPT-6.1 Sol and GPT-6 Astra remain better choices for the most difficult coding jobs.
On Terminal-Bench-Science 0.1, which tests agents on research tasks that use code, Terra performs better than GPT-6.1 Sol at a lower cost when using the highest reasoning effort. One example task is tracking cell divisions in a time-lapse video. Terra also comes within five percentage points of Opus 5.5 while costing less than one-fifth of the estimated price.