Verified Run by Hars @thisishars through Codex, locked to the cube tools. Finished 11 Oct 2026 23:53 UTC.
GPT-6-Astra scores 36.5 at high, well below its best max result of 87.5. Its max result puts it 2nd of 17 models on the Index; this high run is not the one it is ranked on. At level 6 of the 3×3 interactive ladder, it solved the cube in 64 moves, 58 more than the 6-move scramble.
It cleared the 2×2 ladder without a miss, keeping all 3 lives, though its 11 solves used 267 moves. On the 3×3 interactive ladder, misses at levels 4 and 5 left one life for that long level 6 solve; a miss at level 7 ended the round. One-shot, with one answer per cube and no look at the result, stopped after three misses at level 6. It took longer than either interactive round and left the Gauntlet locked.
We spent 47 min 11 s on the high suite, at an API-equivalent cost of $12.48. The medium run took 19 min 26 s and cost $2.84; xhigh scored higher for $18.20 in 1 h 30 min 32 s.
Cleared all 11 levels with all 3 lives left, using 267 moves in 7 min 46 s. The first 2 cubes fell in exact inverses, but level 3 took 118 s and 28 moves, the longest time on any cube. Every later solve also exceeded the scramble length, though the final cube took just 19 s: 28 moves against an 11-move scramble.
At high, it solved through level 6 of 20 before losing its last life on level 7. The level 6 solve took 64 moves against a 6-move scramble, yet finished in 95 s. Levels 4 and 5 took most of the round’s 18 min 22 s, including a 58-move solve of a 5-move scramble. All three misses came from reaching the move limit.
It reached level 5 of 20 before losing all 3 lives to wrong answers at level 6. Every successful answer exactly reversed the scramble, with the first 3 cubes each solved in 15 s or less. Levels 4 and 5 each took more than 3 min, and the longest wait was 442 s for the first wrong answer. At max, the same model cleared all 20 levels on this test.