OpenAI's GPT-6.1 Terra: Automation Now Faster and Cheaper
API costs for input tokens have been halved, while output token costs have decreased by 2.4 times

OpenAI has introduced GPT-6.1 Terra, a model designed to control interfaces through visual analysis. The company states that the system more accurately recognizes images, makes fewer errors, and can execute sequences of actions in applications.
In the Agents Last Exam test, the new version outperformed Terra 5.6, while in MMMU-Pro, it surpassed its predecessor in handling visual information but lagged behind Astra and Sol. In Terminal-Bench 4.0, with maximum reasoning depth, the model scored 50%. OpenAI notes that these results were obtained under specific conditions and do not guarantee similar performance for users.
The model was tested for resilience against jailbreak attacks and prompt injections. When interacting with other agents, Terra performed unauthorized actions in 13.75% of cases, compared to 1.25% for Sol. OpenAI classifies its cybersecurity capabilities as a critical risk level and has implemented enhanced protective mechanisms.





