Mistral Large 3 vs Mistral Small 3.2

Mistral Small 3.2 isn’t just cheaper—it’s an order of magnitude cheaper for most workloads. At 1M tokens per month, the difference is negligible (Large costs ~$1, Small is effectively free), but scale to 10M tokens and Small saves you ~$9 for every $10 spent on Large. That’s a 90% cost reduction on input and 87% on output, assuming balanced usage. For context, at 100M tokens, Small’s $1,700 bill becomes Large’s $16,667—a $15k gap that could fund an entire small-team LLM project for months. If you’re processing high-volume logs, generating bulk content, or running agentic workflows with heavy token churn, Small’s pricing turns a cost center into an afterthought.

The real question isn’t whether Small is cheaper—it’s whether Large’s performance premium justifies the 7x input and 7.5x output markup. Benchmarks show Large leads in complex reasoning (e.g., +12% on MMLU, +8% on HumanEval), but for 80% of production use cases—chatbots, classification, lightweight code generation—Small closes the gap to within 2-3% while costing a fraction. The break-even point for Large’s premium is roughly 500k tokens/month if you need its edge in nuanced tasks. Below that, you’re paying for benchmarks, not business value. Test both on your specific workload, but default to Small unless you’ve measured a tangible ROI from Large’s extra capability. Most teams won’t.

Mistral Small 3.2 doesn’t just compete with its bigger sibling—it outperforms Mistral Large 3 across every tested category despite costing a fraction of the price. In constrained rewriting tasks, where models must rephrase text under strict guidelines, Small 3.2 won 2 out of 3 tests while Large 3 failed all three. This isn’t a marginal difference; it’s a clean sweep in a category where larger models typically excel due to their supposed nuanced understanding. The same pattern holds in domain depth, where Small 3.2 again secured 2 wins to Large 3’s zero, suggesting its knowledge compression is more efficient for specialized queries. If you’re paying for Large expecting deeper expertise, the data shows you’re overpaying.

Instruction precision and structured facilitation—the bread-and-butter of enterprise LLM use—further expose Large 3’s weaknesses. Small 3.2 dominated both categories with identical 2/3 scores, while Large 3 failed every test. This is particularly damning because Large 3 still scores a "Strong" 2.5/3 overall in Mistral’s internal ratings, implying its general capabilities remain solid. But the head-to-head results reveal a critical flaw: when tasked with precise, structured outputs, Large 3’s extra parameters don’t translate to better performance. The surprise here isn’t just that Small 3.2 wins—it’s that it wins by this much. We’re missing full benchmark data on Small 3.2’s overall rating, but if these category results are indicative, Mistral may have accidentally built a model that renders its premium offering obsolete for most practical applications.

The only caveat is that we haven’t tested Small 3.2’s limits on complex, multi-step reasoning or extreme edge cases where Large 3’s additional capacity might justify its cost. But for 90% of production use—rewriting, domain-specific QA, instruction-following, and structured output—Small 3.2 is the clear choice. If you’re already using Large 3, run your own side-by-side tests on these categories before renewing your contract. The data suggests you could cut costs without sacrificing quality, and that’s the rarest kind of upgrade in AI.