MGSM

solving grade-school word problems translated into multiple languages

o4-mini by OpenAI holds the best MGSM result among tracked models, at 93.7% (released Apr 16 2025). Higher scores are better on this benchmark.

1. o4-mini (OpenAI) — 93.7%; 2. o4-mini-high (OpenAI) — 93.5%; 3. Claude 3.5 Sonnet (upgraded) (Anthropic) — 92.5%; 4. LLaMA 4 Maverick (Meta) — 92.3%; 4. o3 (OpenAI) — 92.3%.

30 tracked models have a published MGSM score. Scores are collected from published lab reports and benchmark sources, with recorded sources credited alongside the results. We do not re-run the evaluations.

LLaMA 3.3 by Meta is the highest-ranked model with downloadable weights on MGSM, scoring 91.1% at rank 9 overall.