Notes: GPT-5.6-Luna · Cube Arena

GPT-5.6-Luna gave up nine times with lives left

GPT-5.6-Luna sits 12th on the Index, with its best result at max, but giving up with lives left persisted from low through its highest effort.

Where it stands

GPT-5.6-Luna scores 19.3 at max and sits 12th of 17 models on the Index. Just above it is GPT-5.6-Terra at max, scoring 22.3.

The max run behind GPT-5.6-Luna’s standing reached level 4 of 11 on the 2×2 and level 4 of 20 on both 3×3 tests. GPT-5.6-Terra reached the same point on the 2×2 but went a level further on each 3×3 test. Across GPT-5.6-Luna’s nine suites, none cleared a ladder or unlocked the 4×4 Gauntlet. Its other max run reached level 5 on the 3×3 one-shot test, where each cube gets one answer without seeing the result.

What stood out

We saw GPT-5.6-Luna give up with lives left nine times: three at low, two at high, three at xhigh and one at max. At low, it ended both interactive ladders on level 3 without losing a life. At max, it solved the first four one-shot cubes, then gave up on level 5 with all 3 lives intact. "Test ended at level 4 of 20 (level 5 unsolved)." That round lasted 44 min 54 s and cost $0.35.

The 3×3 cube at level 5, as the model got it — Level 5, as the model got it. Replay

It also threw away eight cubes by repeating a move until the move limit ran out. Each throwaway cost a life; while lives remained, it brought a fresh cube at the same level. At low, it discarded all three level-3 interactive 3×3 cubes, including one with R repeated 26 times. At medium, two 2×2 throwaways included U repeated 48 times on level 4. At high, it repeated R 31 times on a level-3 3×3. At xhigh, it discarded two more 2×2 cubes, including 38 repetitions of U on level 4.

move 0 / 48

Level 4, 48 moves, not solved. Full replay

Effort by effort

low, two runs, 8.1 and 6.3: Neither went beyond level 2. Both exhausted their lives on one-shot after solving only level 1; both gave up on the 2×2.
medium, 10.0: Reached level 3 on the 2×2, but the 3×3 results stayed at level 2 interactive and level 1 one-shot. All three rounds ended out of lives.
high, 14.5: Reached level 3 across all three tests. The 2×2 ended out of lives; both 3×3 rounds ended when it gave up with 1 life left.
xhigh, three runs, 17.5, 14.5 and 13.0: All stopped at level 3 on the 2×2 and gave up on one-shot. One matched high’s score for $0.85 rather than $0.31; another scored lower.
max, two runs, 19.3 and 20.8: Both exhausted their lives after level 4 on the interactive tests. One-shot ended after level 4 by giving up in one run, and after level 5 out of lives in the other.

GPT-5.6-Luna, Fast v3 score at each effort.

Time and cost

The max run behind GPT-5.6-Luna’s Index standing took 1 h 52 min 50 s and cost $1.34 at API-equivalent prices. Across all efforts, suites ranged from 75 s to 3 h 37 min 30 s, and from $0.03 to $1.72. The strongest xhigh run matched that max run’s two 3×3 results while finishing one level lower on the 2×2, for 44 min 4 s and $0.71. The longer max run spent 2 h 28 min 11 s on one-shot alone, ending after three misses on level 6.