Clairic Is Cloudflare's Bet That Extraction Deserves Its Own Model

Cloudflare announced Clairic, an extraction model it calls System One-class, and says it processes unstructured data faster, more cheaply, and more accurately than a normal LLM. Every one of those comparisons is Cloudflare’s own, and the announcement gives me nothing I can score independently.

Extraction is the job most offices actually have: invoices, receipts, shipping forms, the PDF that turns out to be a scan of a fax. It is also where general chat models run up the biggest bill relative to what they return. A model built to emit fields and nothing else is a sensible product.

The accuracy claim is the one to hold loosely. Extraction accuracy depends mostly on the documents themselves, and a narrow model tuned on clean invoices can beat a frontier LLM there and still fall apart on a folder of crooked phone photos.

What I want is the failure list, not the benchmark table: which document types Clairic refuses, and how it reports a field it cannot read. Until someone publishes that, the honest test is the ugliest folder in the office — run Clairic and the current pipeline over the same scans, and count the fields each one gets wrong.