OpenAI GPT-6.1 Terra ReleaseProvider report
2026-10-26Six GPT-6.1 Terra release metrics, with same-chart GPT-6 Astra comparisons for four new metrics, showing the best published reasoning-effort result (minimum for error rates). All five efforts from low to max and per-task USD costs are retained. Agents' Last Exam uses v1; MMMU-Pro uses Vision; OSWorld 2.1 uses the offline set without specifying partial/strict scoring and is kept separate. The difficult-prompt factuality evaluation was refreshed and is separate from earlier GPT-6 reports; error rates do not represent typical usage. These are vendor evaluations; model metadata and historical reports are unchanged.