Composer 2.5 went furthest without feedback
Composer 2.5 reached level 2 one-shot but stopped at level 1 on both interactive ladders, placing it 18th on the Index.
Where it stands
Composer 2.5 scores 6.3 from its only run, sitting 18th among 16 models on the Index. Just above it is GPT-6-Luna at max, with 11.5.
It finished behind that model on both interactive ladders, but went one level further on the 3×3 one-shot test. Neither model unlocked the 4×4 Gauntlet.
What stood out
We saw all 6 interactive misses come on cubes scrambled by just 2 moves, despite the model being able to turn each cube and see the result. Across those rounds, it made 150 moves for 2 solves. On its second 2×2 miss, it said, "Exact original scramble restored." It then announced a try with U2 F2, but the attempt failed.
The one-shot round, with one answer per cube and no look at the result, went further with fewer moves. Composer 2.5 submitted 11 moves across 5 cubes and solved 2. Its misses were split between levels 2 and 3, rather than all falling at the first 2-move scrambles.
The run
The 2×2 opened with a 1-move solve, then ended after 3 failed attempts at level 2, leaving its best at level 1 of 11. The interactive 3×3 followed the same pattern, stopping at level 1 of 20. One-shot survived a level 2 miss before passing that level; both remaining lives went at level 3. Every round ended out of lives, rather than with a give-up.
Time and cost
The suite took 31 min 27 s. The interactive rounds accounted for 26 min 55 s; one-shot took 4 min 32 s. There are no other efforts to compare, and no API-equivalent cost is available. Total output was 51,997 tokens.