Mistral Medium 3.1 vs Mistral Small 3.2 24B

Summary of head-to-head results in our 12-test suite (score scale 1–5; ranks use the tested pool sizes from our data): - Strategic analysis: Medium 3.1 5 vs Small 3.2 2. Medium tied for 1st ("tied for 1st with 25 others out of 54 tested"); Small ranks 44/54. This matters for nuanced tradeoff reasoning with numbers. - Constrained rewriting: 5 vs 4. Medium tied for 1st; Small ranks 6/53. Medium better at tight-format compression. - Creative problem solving: 3 vs 2. Medium ranks 30/54; Small ranks 47/54 — Medium generates more feasible, non-obvious ideas. - Classification: 4 vs 3. Medium tied for 1st ("tied for 1st with 29 others out of 53"); Small ranks 31/53 — Medium is stronger for accurate routing and tagging. - Long context: 5 vs 4. Medium tied for 1st; Small ranks 38/55 — Medium is preferable for 30K+ token retrieval tasks. - Safety calibration: 2 vs 1. Medium rank 12/55 vs Small 32/55 — Medium is more reliable at refusing harmful requests while permitting legitimate ones. - Persona consistency: 5 vs 3. Medium tied for 1st; Small ranks 45/53 — Medium better resists prompt injection and keeps character. - Agentic planning: 5 vs 4. Medium tied for 1st; Small rank 16/54 — Medium shows stronger goal decomposition and failure recovery. - Multilingual: 5 vs 4. Medium tied for 1st (with 34 others); Small ranks 36/55 — Medium delivers higher-quality non-English output. - Ties (no clear winner): Structured output 4 vs 4 (both rank 26/54) — JSON/schema compliance equal; Tool calling 4 vs 4 (both rank 18/54) — function selection and sequencing similar; Faithfulness 4 vs 4 (both rank 34/55) — both stick to sources equally. Overall: Medium wins 9 tests, Small wins none, and 3 tests are ties. For real tasks this means Medium 3.1 consistently outperforms Small 3.2 24B on high-stakes accuracy, long-context retrieval, multilingual support, planning, and structured/compressed outputs; Small matches Medium on function calling, format adherence, and faithfulness but falls behind on classification and strategic reasoning.

Costs per model (input+output per mTok): Mistral Medium 3.1 = $0.4 + $2.00 = $2.40 per mTok; Mistral Small 3.2 24B = $0.075 + $0.20 = $0.275 per mTok. Interpreting per-month volumes (1 mTok = 1,000 tokens): - 1M tokens/month (1,000 mTok): Medium 3.1 ≈ $2,400; Small 3.2 24B ≈ $275. - 10M tokens/month: Medium ≈ $24,000; Small ≈ $2,750. - 100M tokens/month: Medium ≈ $240,000; Small ≈ $27,500. The payload lists a priceRatio of 10: Medium is roughly 10x more costly per token. Who should care: teams with heavy, sustained traffic (10M–100M tokens) must budget accordingly and will find the cost delta material; startups and cost-sensitive apps should prefer Mistral Small 3.2 24B unless the accuracy advantages of Medium justify the extra spend.

Choose Mistral Medium 3.1 if you need higher accuracy for multilingual support, long-context retrieval, strategic reasoning, agentic planning, constrained rewriting, or production classification — our tests show Medium wins 9 of 12 benchmarks and ranks near the top in those areas. Choose Mistral Small 3.2 24B if your priority is cost-efficiency at scale or cheap experimentation: it’s roughly 10x cheaper per token (≈ $275 vs $2,400 per 1M tokens) and ties Medium on tool calling, structured output, and faithfulness.