Opus 5 hits 30% on ARC-AGI-3 benchmark

Opus 5 reaching 30% on ARC-AGI-3 marks a notable leap in novel problem-solving capability, though questions persist around how benchmark evolution impacts out-of-distribution evaluation.

• –Opus 5 achieves a landmark 30% score on ARC-AGI-3, demonstrating measurable progress in zero-shot pattern abstraction.
• –ARC-AGI benchmarks target core reasoning performance where standard pre-training scaling typically yields diminishing returns.
• –Researchers debate whether performance on newer ARC iterations benefits from models being optimized against the paradigms established in ARC 1 and 2.