Claude Opus 5 reportedly gets all 1,200 multiplication problems right without tools at maximum reasoning

A user reported on September 16 that Claude Opus 5 correctly answered all 1,200 multiplication problems involving numbers up to 20 digits each without external tools. They said this required maximum reasoning. For two 100-digit numbers, they reported about 80% accuracy and suggested the drop likely came from exceeding a 128,000-token limit.

The same user later reported roughly 0% accuracy for Fable 5.1 on 5-digit-by-6-digit multiplication at low reasoning. They attributed the failures to skipped reasoning, saying the model answered larger problems correctly when it began producing chain-of-thought reasoning, before accuracy declined again near 20-digit-by-20-digit multiplication.

A reply suggested an external classifier might decide whether to reason, rather than the model itself. The tester countered with a quotation from Claude documentation saying the model decides whether and how much to think.