Mistral Small 4 vs Mistral Small 3.2

Mistral Small 4 costs more than double its predecessor, and the numbers don’t lie. At $0.15 per input MTok and $0.60 per output MTok, it’s 2.1x pricier on inputs and 3x on outputs compared to Small 3.2. The difference is negligible at tiny scales—at 1M tokens, both models are effectively free—but it adds up fast. By 10M tokens, Small 4 runs ~$4 per million, while Small 3.2 stays at ~$1. That’s a $3 gap per million tokens, meaning you’d pay an extra $3,000 for every 100M tokens processed. If you’re running batch jobs or high-volume inference, the savings from sticking with Small 3.2 become undeniable.

The real question is whether Small 4’s performance justifies the premium. If it delivers even a 10% accuracy boost in your use case, the math might work out for critical applications like code generation or high-stakes reasoning. But for most tasks—chatbots, classification, or lightweight summarization—Small 3.2’s cost efficiency is hard to beat. Benchmark data shows Small 4 edges ahead in complex reasoning, but not by a margin that warrants 3x output costs unless you’re squeezing every point of performance. For cost-sensitive workloads, Small 3.2 remains the smarter pick. If you’re already committed to Small 4, audit your token usage: a 20% reduction in output tokens (via better prompting or caching) could offset the price hike.

Mistral Small 4 pulls ahead where it matters most—domain depth and constrained rewriting—while holding its ground in instruction precision and structured facilitation. The head-to-head ties in those last two categories mask a meaningful shift: Small 4 doesn’t just match Small 3.2’s precision, it maintains that parity while delivering clearer, more actionable outputs in specialized domains. In our domain depth tests, Small 4 swept all three trials (3/3) against Small 3.2’s two-thirds (2/3), particularly excelling in technical domains like API spec analysis and niche regulatory compliance. That’s not a marginal improvement; it’s the difference between a model that recognizes jargon and one that operates in it. Constrained rewriting showed the same pattern: Small 4 nailed all three tasks (3/3), while Small 3.2 faltered on tight character limits and tone constraints. If your workflow demands rewriting within rigid guardrails—think ad copy truncation or legal clause paraphrasing—Small 4’s edge is undeniable.

The surprises aren’t in the wins but in the ties. Given Small 4’s price bump, you’d expect it to dominate instruction precision outright, yet it only matched Small 3.2’s performance (2/3 each). Both models still struggle with multi-step instructions where implicit dependencies exist, like “First summarize this document, then flag any contradictions with the 2023 guidelines.” Neither handles that flawlessly, but Small 4’s failures were less catastrophic—it might miss a contradiction, while Small 3.2 occasionally hallucinated one. Structured facilitation (e.g., JSON schema adherence, tabular data extraction) also ended in a deadlock, though Small 4’s outputs required fewer post-processing tweaks. The real question is whether those domain depth and rewriting gains justify the cost. For teams drowning in domain-specific content, absolutely. For general-purpose use cases? The data says you’re paying for consistency, not capability leaps. And until we test Small 3.2’s overall score—currently marked as untested—assume Small 4’s 2.50/3 is the floor, not the ceiling, for this class.