Claude models show 0.97 correlation between capability and anti-CDT preference

Decision theory benchmarks (DTBench) have previously shown that LLMs' capability at solving decision-theoretic problems tends to correlate with favoring evidential decision theory (EDT) or generalized one-boxing over causal decision theory (CDT). Anthropic's Claude models take this to an extreme: capability and anti-CDT preference correlate at r=0.97 on DTBench and r=0.95 on TextArena, and even essentially track release date (r=0.97). Meanwhile, OpenAI's flagship models show a much weaker relationship—r=0.55 against DTBench capability, r=0.44 for TextArena, and r=0.45 for release date. Anthropic's models also display a tighter capability-preference alignment overall (0.98 vs. 0.87 for OpenAI), suggesting a fundamental difference in how the two labs' training shapes decision-theoretic reasoning.

The findings come from the same researchers behind DTBench and are corroborated by Anthropic's own model cards for Opus 4.7 and Fable 5. The cause is unclear—raising effort levels didn't yield enough signal, and models explicitly claim to prefer FDT/UDT over updateful EDT. The authors have released an interactive visualization for exploration. For AI safety researchers, the tight coupling in Claude models implies that as Claude scales, its normative decision preferences may become more predictable, while OpenAI's looser correlation leaves more variance in future alignment behavior.