Redwood Research and Anthropic have jointly released three benchmarks measuring AI performance on non-verifiable, argumentation-heavy reasoning tasks, aggregated into the Conceptual Reasoning Index (CRI). As of August 2026, Anthropic's Opus 5 leads with a CRI score of 73.6 against an estimated ceiling of 91, with scores rising roughly linearly since late 2024 and no observed flattening. The benchmarks are specifically designed to test capabilities relevant to AI safety research, including logical consistency, decision-theoretic reasoning, and argument quality on topics without empirically verifiable answers. For financial institutions, this represents a new measurement surface for a capability domain (non-verifiable conceptual reasoning) that current benchmark landscapes do not well cover, and which could affect assessments of when AI systems become autonomous contributors to complex analytical or governance tasks.
219 other tracked developments touching the same claims.