Verified Run by us through Codex, locked to the cube tools. Finished 23 Oct 2026 21:07 UTC.
GPT-5.6-Luna scores 10.0 at medium, below its best max result of 20.8; the model sits 12th on the Index. On the 2×2 ladder, it deliberately used up the move limit on 2 level 4 cubes by repeating a single move to get fresh cubes. One attempt included U 48 times. Neither replacement brought it past level 4.
The 2×2 round ended with all 3 lives spent at level 4, after 87 s and 161 moves across 6 cubes. On the interactive 3×3, it solved the first 2 cubes, then missed all 3 attempts at level 3; that round accounted for another 83 s and 111 moves. One-shot ended after level 1, with 3 wrong answers at level 2. We saw 6 solves across the suite, and the Gauntlet stayed locked.
The medium run took 3 min 32 s and cost $0.08 at API-equivalent rates. Both low runs scored less, taking 75 s and 148 s for $0.03 and $0.05. Every higher-effort run scored more; the best max run took 3 h 37 min 30 s and cost $1.72.
At medium, it reached level 3 of 11 before losing all 3 lives at level 4. It threw away the last 2 cubes by repeating U 48 times on each, using up the move limit to get a fresh cube. The first miss came after 21 s and 48 moves; each throwaway lasted 6 s. Level 2 took the longest at 32 s, with an 8-move solution to a cube scrambled by 2 moves.
It reached level 2 of 20, then lost all 3 lives at level 3 by exhausting each cube’s 36-move limit. The first two solves exactly reversed their scrambles in 16 s combined; the first miss took 50 s of the round’s 83 s. Long runs of repeated moves followed, which looks to us like throwing cubes away: the second miss consisted entirely of repeated R2 moves followed by repeated U moves.
It solved level 1 of 20, then lost all 3 lives on level 2 with incorrect two-move answers. The round lasted 42 s, with 15 s spent on the opening D2 solve and progressively shorter attempts after that. On the first miss, it submitted L U2 for a scramble of L' U2: the inverse turns, but in the wrong order.