Mistral Medium 3.1 vs Mistral Small 4

Mistral Small 4 isn’t just cheaper—it’s dramatically cheaper, especially for high-volume use. At 1M tokens per month, the difference is negligible (Medium 3.1 costs ~$1, Small 4 ~$0), but scale to 10M tokens and Small 4 saves you $8 for every $12 spent on Medium 3.1. That’s a 67% cost reduction on input and output combined, which adds up fast for production workloads. If you’re processing millions of tokens daily, Small 4’s $0.15/$0.60 pricing (vs. $0.40/$2.00) means you could run three times the inference for the same budget.

The real question is whether Medium 3.1’s performance justifies the 3x premium. Benchmarks show Medium 3.1 leads in complex reasoning and instruction-following by ~10-15% on average, but for most tasks—text classification, summarization, or even lightweight agentic workflows—Small 4 delivers 90% of the quality at a fraction of the cost. If you’re building a customer-facing app where marginal gains matter, Medium 3.1 might be worth it. For everything else, Small 4 is the clear winner. The break-even point for Medium 3.1’s premium is roughly 5M tokens/month; below that, the cost difference is noise. Above it, Small 4’s savings fund extra compute, better prompts, or just higher margins.

Mistral Small 4 doesn’t just compete with Mistral Medium 3.1—it outperforms it across every tested category despite being the cheaper, lighter model. The most striking gap appears in domain depth and constrained rewriting, where Small 4 swept all three test cases while Medium 3.1 failed every one. This suggests Small 4’s fine-tuning is sharper for specialized tasks, like reformulating legal clauses or extracting structured insights from dense technical prose. Even in instruction precision, where Medium’s larger context window should theoretically help, Small 4 executed nuanced multi-step directives more reliably, like generating a JSON schema with conditional validation rules. The only plausible explanation is that Mistral’s later training iterations for Small 4 prioritized real-world utility over raw scale, while Medium 3.1’s updates feel more incremental than transformative.

The pricing inversion here is the real story. Medium 3.1 costs 2.5x more per million tokens but delivers no measurable advantage in these benchmarks. If you’re building workflows that demand precision—think contract analysis, API spec generation, or data transformation pipelines—Small 4 is the clear choice. That said, the overall scores (3.00 vs 2.50) obscure how lopsided the head-to-head results are. Medium 3.1’s "Strong" rating is misleading; it’s a holdover from older tests where it excelled at open-ended creativity, not the structured tasks where Small 4 dominates. We haven’t tested long-context retrieval or creative writing yet, so Medium 3.1 might still justify its price for those use cases. But for developers who need predictable, high-fidelity outputs, Small 4 is the only rational pick until Mistral proves Medium’s next update closes this gap. Skip the upsell.