Anthropic's Opus 5 blows past Fable 5 and GPT-5.6 Sol on the benchmark designed to measure real intelligence

Anthropic's Claude Opus 5 achieved a 30.2 percent score on the ARC-AGI-3 benchmark, nearly quadrupling the previous record of 7.8 percent set by OpenAI's GPT-5.6 Sol. The benchmark's developers report that Opus 5 independently formulated reflection equations, a capability they had not observed in any other model. The developers attribute this breakthrough to stronger logical reasoning in the architecture.

ARC-AGI-3 measures abstract reasoning and general intelligence through tasks designed to test problem-solving beyond pattern matching. The benchmark strips away language-specific knowledge and tests how models handle novel reasoning challenges. A score of 30.2 percent represents a substantial leap in model performance on tasks that require genuine analytical thinking rather than memorized patterns.

The reflection equations observation carries particular weight. When a model independently generates these equations without explicit instruction, it signals deeper structural understanding rather than surface-level pattern recognition. This behavior suggests Opus 5 possesses reasoning capabilities that go beyond statistical associations in training data.

Fable 5, mentioned in the headline, received no specific score in the available details, but fell behind both Opus 5 and the previous GPT-5.6 Sol record. The dramatic performance gap between Opus 5 and the former leader raises questions about architectural differences and training methodologies between Anthropic and competitors.

ARC-AGI benchmarks have gained credibility among researchers as meaningful measures of model reasoning because they resist brute-force scaling and require actual problem-solving logic. Achieving such a significant performance increase on this particular metric matters more than improvements on traditional benchmarks that models can partially solve through pattern memorization.

The result demonstrates that reasoning ability, rather than raw scale or parameter count, differentiates advanced models. Opus 5's independent formulation of reflection equations without