THE CRUNCH
Cloudflare has announced a preview of Clairic, an extraction model hosted on Workers AI as @cf/cloudflare/clairic, with open weights planned for release soon. The company positions it as the first of a new class of "System One" extraction models: instead of generating text token by token like a large language model, it answers every requested field in a single pass, returning typed values such as names, dates, numbers and lists, each with a confidence score. If a value is missing, it returns null rather than inventing one.
Users name the fields they want, such as an order number or invoice total, and Clairic returns matching values in the right format. Beyond straightforward extraction, it can compute values from the input, estimate figures the input only implies, and forecast future values of a series, with estimates returned as distributions and an 80% interval. Because it generates no tokens, Cloudflare says it cannot produce a mistyped answer or hallucinate in the way an LLM can, though it can still make mistakes.
The model follows TypeSafe's Jev, a "System One" decision model introduced about a month earlier that returns probabilities against a fixed list of questions. Clairic complements that: decision models turn typed values into choices, while extraction models turn raw input into those values. Cloudflare suggests chaining them with an LLM, giving the example of a complaint email where Clairic pulls out the order number, amount and damage and estimates repair cost, the decision model Clef picks between a refund and a repair, and GLM-5.3 writes the reply.
Cloudflare says that chain takes 19.3 seconds, against 61.7 seconds and nearly double the cost if GLM-5.3 handles all three steps alone. There is a limit, though: the company says Clairic cannot yet reliably handle longer chains of calculations, so developers should run those in their own code for now.
IN NUMBERS