Claude Opus 5 resets the AI benchmark ceiling

Anthropic's Claude Opus 5 launched today and landed at the top of the Artificial Analysis Intelligence Leaderboard, posting 30.2% on ARC-AGI-3, a meaningful gap above the next best model. The threads on both the Opus 5 announcement and the leaderboard discussion were immediately dominated by two questions: what does this cost per token, and does the benchmark performance translate to real work?

The pattern here is the gap between benchmark performance and felt experience. Multiple commenters noted they felt like they were back to Opus 4.5-level results in daily use after a few weeks, regardless of which model they used. One commenter flagged that Fable, a model some feel is the real ceiling, isn't even on the ARC-AGI leaderboard, which raises questions about what these rankings are actually measuring. The 'AA-Omniscience Index' measuring hallucination was also called out as an interesting and underappreciated signal.

The counterpoint worth noting: cost-per-performance is genuinely improving. The discussion acknowledged that intelligence vs. output tokens per dollar is now a real axis to optimize on, not just raw capability. Founders building on top of these models should treat that cost curve as a business variable, not a fixed constraint.