Claude Opus 5.5 Wins Six AAArena Golds

This is a strong model-plus-harness result, not proof of universal human-level performance. The six unclaimed ladders show that strategic learning still breaks down as game rules and planning demands grow.

• –AAArena uses 12 adversarial games and 1,920 archived human programs, creating more realistic baselines than synthetic puzzle evaluations.
• –Fixed weights isolate the value of replay analysis, opponent selection, feedback loops, and executable policy revision.
• –Claude Code is part of the measured configuration, reinforcing that developers should evaluate complete agent systems—not model names alone.
• –Six gold medals indicate broad capability, but failing to top the remaining ladders exposes clear limits in complex, long-horizon reasoning.