OpenAI introduces GPT-6.1 Terra, a new model focused on speed and visual reasoning. The model is designed to carry out complex computer tasks and writing work more efficiently and at lower cost.
OpenAI has announced a brand-new AI model. GPT-6.1 Terra focuses on carrying out ‘ordinary’ tasks efficiently, such as drafting texts, interpreting images and filling in forms. The model tries to strike a balance between speed, capacity and the cost to the user. The Terra models are comparable to Claude Haiku and Gemini Flash.
Improved visual tasks and computer use
According to benchmarks shared by OpenAI, the new model excels at tasks in which the AI interacts directly with a computer environment. In the Agents’ Last Exam benchmark, in which AI agents carry out professional tasks such as converting architectural drawings into 3D models, the model scores considerably higher than its predecessor. Be critical of benchmarks: OpenAI only shows the figures that fit the picture it wants to paint.
At medium effort, GPT-6.1 Terra scores 15 percentage points higher than GPT-5.6 Terra, at roughly one third of the cost of the previous model. The model also performs better than the previous version when using desktop applications, as measured by OSWorld 2.1.
Besides computer use, there is also progress in visual reasoning. On the MMMU-Pro benchmark, the model comes close to the scores of the most powerful models in the GPT-6 family.
Fewer errors and better code
OpenAI has invested heavily in reducing hallucinations and factual errors. Internal benchmarks show that the error rate at low effort has fallen from 53.6 percent to 16.7 percent. At maximum effort, the error rate drops even further, to 4.3 percent.
The leap is also big for software development. In Terminal-Bench 4.0, which measures how well AI agents solve problems via the command line, GPT-6.1 Terra solves more than 50 percent of the tasks. That is more than double the score of GPT-5.6 Terra, which stood at 21.5 percent.
The model also performs well in scientific research via code, such as tracking cell divisions in videos. In Terminal-Bench-Science 0.1, it comes close to the performance of more expensive competitors.
Availability and safety
On safety, the model meets the critical threshold for cybersecurity under OpenAI’s framework. Under the agreement with the US government, which is only one page long, AI companies determine that threshold entirely themselves. The model is also better able to communicate honestly about its own limitations and resists malicious instructions via prompt injection.
As of today, GPT-6.1 Terra is available to users of ChatGPT Work and Codex via the Plus, Pro, Enterprise and education plans. Developers can use the model via the OpenAI API at a price of 1 dollar per million input tokens. GPT-6.1 Terra follows GPT-6.1 Sol and GPT-6 Astra.