Frontier LLMs Nail 20-Digit Multiplication Without Tools

On September 17, developer maksymandr published a series of evaluations showing that frontier LLMs can multiply two 20-digit numbers with near-perfect accuracy without calling any external tools, far exceeding human mental arithmetic.

Confirmed

• Claude Opus 5 answered all 1200 problems in a 20×20-digit multiplication test correctly (100%) at maximum reasoning effort, though the result depends on the highest reasoning tier (m1, m7).
• GPT-6-Astra cannot run in a no-CoT mode — its lowest API reasoning tier is low — and at that tier its multiplication accuracy is still 99.6% (m3).
• cozyblazex replicated the test with GPT-5.5 at medium reasoning effort and 7 samples per cell, achieving 99.46% on multi-digit multiplication with no tools invoked; maksymandr cited the result and added practical advice (m6).
• maksymandr clarified the answer-only prompt setup: the model must output only the final integer answer with no commas, text, or intermediate steps, meaning all computation must happen within a single (possibly looped) forward pass (m9).
• He also posted a long-multiplication worked example showing this scale is virtually impossible for humans by hand (m2).
• Practical note: any reasonable LLM should call an external tool for multiplication beyond 5×5 digits, which guarantees 100% correctness (m4).

Speculation

• On why Fable and Astra lack a non-thinking mode, maksymandr speculates: it would let others infer model size and depth, no-CoT versions are more attack-prone, and only Fable can read Fable's hidden chain of thought, a channel the maker may not want bypassed (m5). This is unofficial speculation.
• scaling01 notes the benchmark is useful for measuring no-CoT capability, especially for recurrent-depth architectures (m8).

Why it matters

The benchmark quantifies frontier models' pure-reasoning ceiling with a minimal, reproducible task, replicated across models and people with near-perfect results. The strict answer-only setting implies exact arithmetic far beyond human capability within a single forward pass, and it also surfaces a trend of the newest generation of models deliberately dropping "no-thinking" modes, raising questions about the trade-off between capability disclosure and security.

2026-09-17 ~ 2026-09-17 · 9 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
• LLMs Should Just Use Tools: Any Reasonable Model Nails 20x20 Multiplication — maksym_andr · 2026-09-17
• [source] 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
• No chain-of-thought allowed: probing models forced to compute in a single forward pass — maksym_andr · 2026-09-17

1 near-duplicate retellings: xeophon