The key facts

  • OpenAI released the GPT-6.1 Terra model with an error rate of 4.3% at maximum reasoning levels.
  • The new model is 3.5 times cheaper than the previous version when used by AI agents to complete tasks.
  • Access to the model is open for Plus, Pro, Enterprise, and Edu subscribers, as well as developers via API.

OpenAI's new GPT-6.1 Terra model

OpenAI has unveiled a new version of its mid-tier artificial intelligence model, GPT-6.1 Terra, which combines enhanced visual reasoning, precise computer control, and a more natural writing style. According to the organization's press office, the key achievement is a significant reduction in error rates: in internal tests at low reasoning levels, the rate dropped from 53.6% in the previous GPT-5.6 Terra version to 16.7%, while at maximum reasoning levels, errors were reduced to just 4.3%.

Efficiency and cost of use

In the Agents' Last Exam test measuring AI agents' ability to perform complex tasks, the new model scored 60% at a cost of $1.01, demonstrating improved efficiency compared to the previous version, which achieved 50.7% at a cost of $3.58. For comparison, Anthropic's Claude Opus 5.5 model reached a score of 63.2% but required expenditures of $8.82. Access to GPT-6.1 Terra is open for Plus, Pro, Enterprise, and Edu subscribers on ChatGPT Work or Codex services, as well as for developers via API, where the cost is $1 per 1 million input tokens and $5 per 1 million output tokens.