how opus 5 obliterated the arc-agi-3 record using raw algebra

anthropic's claude opus 5 scored 30.2% on the arc-agi-3 benchmark, crushing openai's gpt-5.6 sol record of 7.8%. the model displayed novel behaviors like translating tasks into algebraic notation and formulating reflection equations to solve five previously unsolved environments. arc prize attributes this lead to genuinely better logical reasoning, enabling autonomous planning in unfamiliar environments. however, independent tests on guanghan ning's witness benchmark show opus 5 only tying with competitors like fable 5, pointing to narrower genre-specific gains. developers should monitor whether this raw reasoning holds up in production or simply reflects targeted reinforcement learning.