GPT-6 Astra makes a massive leap on ZeroBench (an extremely difficult vision benchmark), surpassing the human baseline across all three metrics
retained text
pass@5: Scores if at least one of the 5 attempts is correct pass\^5: Scores only if all 5 attempts are correct (a reliability metric) "pass@1": Not a true single-attempt pass@1, it's the average score across 5 attempts. [https://zerobench.github.io/](https://zerobench.github.io/)
document fetched 2026-09-24 00:33 · canonical https://reddit.com/r/singularity/comments/1wo5c7t/gpt6_astra_makes_a_massive_leap_on_zerobench_an