Cloudflare's Clairic takes on LLMs for document data extraction

Cloudflare is pitching Clairic not as a replacement for large language models but as a specialized component in a larger agent chain. In its example, Clairic pulls the order number, amount, and damage from a complaint email and estimates a repair cost; a decision model called Clef applies company rules to choose between a refund and a repair; and GLM-5.3 writes the customer reply. Cloudflare says this chain takes 19.3 seconds, while GLM-5.3 alone needs 61.7 seconds for all three steps and costs almost twice as much.

The company also frames Clairic alongside other "System One" models like Jev, suggesting a division of labor where fast, narrow models handle structured extraction and LLMs handle generation. Cloudflare acknowledges a limit: Clairic cannot yet reliably handle longer chains of calculations, and developers should run those in their own code for now.