Anthropic's Claude models display a near-perfect correlation between raw capability and a preference against CDT (Causal Decision Theory) answers, with r=0.97 when measured by DTBench and r=0.95 by TextArena.
This tight link does not appear in OpenAI models, where the same correlation is only r=0.55 or r=0.44, suggesting Claude's design or training creates a unified scaling law between capability and decision-theoretic reasoning preference that is absent in GPT.
What happened
Researchers found that for Anthropic's Claude models, capability (measured by DTBench or TextArena benchmarks) correlates almost perfectly with a preference against CDT (Causal Decision Theory) answers — with correlation coefficients of r=0.97 and r=0.95 respectively. For Claude flagship models, this correlation is essentially identical to the correlation with release date (r=0.97).
Why it matters
The finding suggests that for Claude, higher capability and philosophical preference for EDT/FDT/UDT over CDT are not separate properties but emerge together as the model scales. This differs sharply from OpenAI models, where the same correlations are much weaker (r=0.55 for capability vs. CDT preference, r=0.44 for TextArena, r=0.45 for release date), indicating Claude's training may systematize decision-theoretic reasoning differently than GPT.
What to watch
Anthropic has already replicated this pattern in their Opus 4.7 and Fable 5 model cards. The correlation holds across different capability measures (DTBench and TextArena), though TextArena and DTBench themselves show a higher correlation within Anthropic models than when all models are pooled.
Ask the AI about this article →
The research documents a striking divergence between how Anthropic and OpenAI models scale decision-theoretic reasoning. For Claude, the near-perfect correlation (r=0.97) between capability and EDT/FDT/UDT preference suggests that as the model becomes more capable, it naturally gravitates toward these decision theories. This is reinforced by the fact that for flagship models, this correlation mirrors the correlation with release date (r=0.97), implying that capability and decision-theoretic preference advance in lockstep as Anthropic releases newer versions.
In contrast, OpenAI models show a weak correlation (r=0.55 or lower) between capability and CDT preference, indicating that raw capability and decision-theoretic leaning are largely independent properties in GPT. This gap points to a possible difference in training objective, alignment approach, or architectural bias. Anthropic's results (including those on Opus 4.7 and Fable 5) confirm the pattern is reproducible within their model family. The finding is further grounded by the observation that TextArena and DTBench themselves correlate more strongly within Anthropic models than across the broader model population, suggesting Anthropic's models behave more consistently along both capability and decision-theory axes.
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytime
Ask AI anything about this article. Q&As are published on this page for other readers too.
Anthropic launched Claude Academy on August 20, a free learning site that explains AI fundamentals and how to…
OpenAI rolled out support on Thursday for controlling Apple's iMessage service via ChatGPT on Mac, enabling th…
As AI technology matures, the bottleneck in the industry is moving beyond semiconductor constraints like GPUs…
OpenAI has launched an Apple Messages plug-in for ChatGPT that lets users connect their Messages inbox to the…
Amazon Bedrock now supports OpenAI GPT-5.6 models (Sol, Terra, and Luna variants) across more than 25 AWS Regi…
Cisco is transforming its digital customer experience (DCX) strategy by embedding AI throughout customer journ…
