Mistral Small: the most boring model I recommend the most
The most expensive habit in production LLM systems is sending every request to a model that could pass a PhD qualifier. Classification, extraction, normalization, dedup, summarize-this-ticket, is-this-spam: the bulk of real traffic is work a strong small model does indistinguishably from a frontier one, at two orders of magnitude less money. I've made this argument abstractly in the cost-architecture post. Mistral Small 26.03 is what the argument looks like as a product.
The spec sheet that matters
Fifteen cents per million input tokens, sixty out. 262k context. Apache-2.0 weights on Hugging Face, so the same model that serves your API tier can run on-prem, air-gapped, or fully local when a customer contract demands it. No SWE-bench run published. For once I don't much care, because nobody hires this tier to close GitHub issues.
What I hire it for, concretely, in systems I run today: routing decisions in the autorouter, first-pass triage of inbound tickets, structured extraction from invoices and logs, subagent scut work where a bigger model orchestrates, and semantic-dedup of a content pipeline. None of that is demo material. All of it used to run on a model costing twenty times more, and when I swapped it out, the eval deltas were inside noise.
The frontier model is for decisions you can't afford to get wrong. Small is for the 80% of tokens where being wrong costs a retry.
The Apache clause is the feature
Plenty of models are cheap. What makes Small strategically interesting is the license. Apache-2.0 means no usage restrictions to lawyer over, no research-only asterisk, no rug-pull risk if the vendor pivots (concerns that killed two otherwise-fine model choices in projects I've consulted on). It also means the cheap tier of your local-first cascade and the cheap tier of your cloud setup can be the same model, which makes eval parity trivial. That's rarer and more valuable than a benchmark point.
Where it breaks
Multi-step reasoning under ambiguity. Ask it to plan, and the plans are plausible and shallow; ask it to follow a crisp instruction against clear input, and it's a metronome. The skill is writing prompts that need no judgement. That's the same discipline, usefully, that makes any pipeline debuggable. It also drifts on very long structured outputs; I cap generations and paginate rather than trust it with a 4,000-token JSON blob.
The routing rule I actually use: if the task has a rubric, Small. If the task is the rubric, frontier. Check the prices page: the gap between those two lines is most of your bill, and closing it is a one-day refactor. Boring is the highest compliment I give infrastructure, and Small 26.03 is extremely boring.