Frontier LLMs Nail 20x20 Digit Multiplication Without Tools

Developer maksymandr posted a series of evaluations and discussions on September 17, with the headline finding: frontier LLMs can already complete 20×20-digit multiplications — far beyond human mental math — with extremely high accuracy, without calling any external tools.

Confirmed

• Claude Opus 5, at its highest reasoning setting, completed a 20×20-digit multiplication test (two 20-digit numbers multiplied) with all 1,200 questions correct — 100% accuracy; however, this depends on maximum reasoning strength being enabled (m1, m6).
• GPT-6-Astra cannot run in a no-CoT mode; the API's lowest reasoning tier is "low," and at that tier its multiplication accuracy was still 99.6% (m3).
• maksymandr posted a long-hand multiplication layout for 20-digit numbers, visually showing that human manual calculation at this scale almost certainly produces errors (m2).
• Practical takeaway: any reasonable LLM should invoke an external tool for multiplications beyond 5×5 digits, which yields 100% correct results (m4).

Views and speculation

• On why the newer models Fable and Astra offer no-thinking modes no longer, maksymandr speculated the reasons include: a no-CoT version would let model size and depth be inferred; no-CoT versions are more vulnerable to various attacks; and only Fable can read Fable's hidden chain of thought, which they don't want bypassed (m5). This is speculation, not officially confirmed.
• scaling01 argued the benchmark can measure no-CoT capability, which is especially meaningful for recurrent-depth architectures (m7).

Why it matters

The evaluation quantifies the ceiling of pure-reasoning capability in frontier models with a minimal, reproducible task, while also revealing a likely deliberate removal of "no-thinking modes" in the latest model generation, sparking discussion of the trade-off between capability disclosure and safety considerations.

2026-09-17 ~ 2026-09-17 · 8 related posts

• [source] Claude Opus 5 aces 20x20-digit multiplication eval, 1200/1200 correct with max reasoning — maksym_andr · 2026-09-17
• [source] GPT-6-Astra reportedly can't run without CoT; even 'low' effort hits 99.6% multiplication accuracy — maksym_andr · 2026-09-17
• What a 20x20-digit multiplication actually looks like by hand, per one LLM eval author — maksym_andr · 2026-09-17
• [source] LLMs Should Just Use Tools: Any Reasonable Model Nails 20x20 Multiplication — maksym_andr · 2026-09-17
• GPT-5.5 hits 99.46% on multi-digit multiplication with pure reasoning, no tools — maksym_andr · 2026-09-17
• Opus 5 aces 20-digit multiplication; speculation on why no-thinking modes are withheld — scaling01 · 2026-09-17
• Why Fable and Astra ship without a non-thinking mode: attack surface and CoT monitoring — maksym_andr · 2026-09-17

1 near-duplicate retellings: xeophon