Verified Run by Rohit @rohit3a through Codex, locked to the cube tools. Finished 11 Oct 2026 23:14 UTC.
GPT-6-Astra at max scores 87.5, its best result and the one that places it 2nd of 17 models on the Index. It cleared all three opening ladders to unlock the Gauntlet, then finished one level behind Claude Opus 5.5 at xhigh on each 4×4 test.
The 2×2 ladder ended without a miss. On 3×3 interactive, it recovered from a miss at level 6 and cleared the ladder with 2 lives left. The one-shot ladder was perfect on results, though level 12 needed 103 moves against a 12-move scramble, 91 more than the scramble. On 4×4 interactive, it gave up at level 8 with 2 lives still available, after an earlier miss at level 6. The one-shot Gauntlet went one level further, then ended with all three misses at level 9.
We measured about 3 h 28 min and $51.94 in API-equivalent cost, compared with about 1 h 31 min and $18.20 at xhigh, which did not unlock the Gauntlet. The 3×3 interactive round accounted for $29.27, more than half this suite’s cost.
It cleared all 11 levels with all 3 lives left, using 119 moves in 20 min 48 s. The first five solutions matched their scramble lengths, but level 5 was the slowest cube at 6 min 32 s. Levels 6 and 8 each took another 4 min 9 s, with solutions of 22 and 21 moves against scrambles of 6 and 8. It finished level 11 in 70 s with 11 moves, as short as the scramble.
It cleared all 20 levels with 2 lives left in about 48 min. The first 3 cubes took solutions as short as their scrambles; the only miss came at level 6, where it exhausted the 72-move limit. Level 4 took the longest: 6 min 23 s for a 44-move solution to a cube scrambled in 4 moves. After solving the replacement at level 6, it finished without another miss, closing with 56 moves on the final cube.
At max, it cleared all 20 levels with all 3 lives left. Level 12 took 103 moves against a 12-move scramble, with no chance to check the result before the answer counted. The first 6 solutions exactly reversed their scrambles; every later cube took at least 53 moves. Level 7 accounted for nearly a third of the round’s time at 12 min 30 s, ending with an 83-move solution, just 1 move below the limit.
At max, it solved through level 7 of 20, then gave up on level 8 with 2 lives left. The first five cubes fell to exact inverses of their scrambles. Its only move-limit miss came at level 6, consuming 72 moves and 18 min 44 s; the replacement took 37 moves, and level 7 took 43, against scrambles of 6 and 7 moves. On the final cube, it spent 11 min 37 s and made just 4 moves before conceding.
Reached level 8 of 20 before three wrong answers at level 9 used all its lives. Every successful answer matched its scramble length, from 1 move through 8 moves. Levels 5 through 7 took more than half the round’s time, with level 6 alone taking about 13 min for a 6-move solution. It finished one level behind Claude Opus 5.5 at xhigh, which reached level 9 on this test.