[X]score: 0.50
Anthropic Opus 4.8 Scores 1.5% on ARC-AGI-3, New SOTA at ~$10K Cost
June 1, 2026
Opus 4.8 achieves 1.5% on ARC-AGI-3 — the current state of the art — at an estimated cost of ~$10K per evaluation run. Qualitative analysis notes the model reasons at a higher abstraction level than Opus 4.7, treating tasks as object-and-system problems, though it still commits to incorrect sub-goals on harder levels.
HOW THIS AFFECTS YOU
●
researcherThe abstraction-level shift between Opus 4.7 and 4.8 on ARC-AGI-3 is a concrete behavioral signal worth analyzing for what architectural or training changes drive higher-order reasoning.