Cloudflare has released Clairic in preview on Workers AI, offering a specialized alternative to general language models for pulling structured fields from documents. Open weights are planned but are not yet available.
A developer specifies fields such as an invoice total or order number. Clairic returns matching values in the requested format, adds a confidence score, and uses null when a value is absent. Because it predicts all requested fields in one pass rather than generating an answer token by token, Cloudflare says it can reduce latency and cost.
The company demonstrated a complaint-processing chain in which Clairic extracts order details and estimates repair cost, a separate decision model applies company rules, and GLM-5.3 writes the response. Cloudflare reports that the chain took 19.3 seconds, compared with 61.7 seconds when GLM-5.3 performed all three jobs, at almost twice the cost.
Those figures come from Cloudflare’s own example rather than an independent benchmark. The company also says Clairic is not reliable enough for long calculation chains, which should still be handled in conventional code.