GPT-6.1 Terra cuts hallucination to 38%, halves price

GPT-6.1 Terra is the first Terra-class model since GPT-5.6, and the pricing shift is steep: half or less per token, with cache writes halved to $1.25 per million tokens and cached input at one quarter of the previous rate. That structure rewards workloads that repeatedly reuse context.

The biggest behavioral change is in AA-Omniscience, where the model now guesses 38% of the time when it lacks an answer, down from 88% for GPT-5.6 Terra. That drop comes alongside a 2 point accuracy gain, so the model is answering more questions correctly while declining more often. On Terminal-Bench 4.0 it gains about 20 points, rivaling GPT-6.1 Sol at max effort.

The cost-efficiency picture is mixed by effort setting. Low and medium effort are Pareto-optimal for Intelligence vs Cost, while max effort sits behind the frontier and scores 4 points below GPT-6.1 Sol but 10 points above GPT-6 Luna on the Intelligence Index. In the Coding Agent Index, the xhigh and low effort settings sit on the cost-efficiency frontier.