Claude Opus 5, handed nothing but stock Claude Code, scored 96.2% on the public ARC-AGI-3 games against 30.2% model-only, for about $540 — building and discarding its own parsers and simulators like a mathematician burning scratch paper.
A neurosurgery resident with no specialized mathematical training cracked the two-decade-old Crouzeix conjecture with a sixteen-hour autonomous run, verified by Crouzeix himself.