Cloudflare has launched Clairic, a model purpose-built for structured extraction - pulling defined fields out of invoices, forms, contracts and similar documents. The company says it runs 3.2x faster and 49% cheaper than doing the same job with an all-LLM pipeline.
The economics are the argument. Extraction is one of the most common production AI workloads and one of the worst matches for a frontier model. The task is narrow and the output is a schema, but most teams reach for a general-purpose model because it is already integrated, then pay general-purpose prices on every document, forever. A dedicated model is not more capable; it is correctly sized, and at volume that is the difference between a line item and a problem.
It fits a clear October pattern. Microsoft-Decision-1 at 9B, Liquid AI's d1-3B and d1-omni-600M, Scira's Amos 1 family from 0.5B, Cloudflare's own Clef models, Fastino GLiDE, Perplexity's pplx-decider - a whole category of small, single-purpose models appeared this month with almost no coverage. The headlines went to frontier releases. The deployments are quietly going somewhere else.