Overview
- OpenAI announced and rolled out GPT-6.1 Terra on Monday, upgrading the previous GPT-5.6 Terra and making the model available through ChatGPT Work, Codex and the API.
- OpenAI says the new Terra is about 50% cheaper than the version it replaces and delivers large internal error-rate improvements, with low-effort errors falling from 53.6% to 16.7% and max-effort errors to about 4.3%.
- OpenAI put Terra into Arena.ai’s Agent Arena and Code Arena so independent testers can run multi-step agentic tasks and coding benchmarks against the model.
- Terra’s public debut matters because the model was previously used internally as a judge in Microsoft coding evaluations, which lets researchers now compare how a former evaluator performs when it is itself tested.
- Analysts say Terra follows Sol in driving down per-task costs for useful AI, a trend that could speed enterprise adoption, squeeze downstream margins and change which tasks companies choose to automate.