Cloudflare has released Clairic, a model designed to pull structured data from unstructured documents faster and more cheaply than large language models. It does this without generating a single token of unnecessary text, which is either a virtue or a personality trait, depending on your perspective.
It is available now in preview on Workers AI. Open weights follow shortly.
A chain of three specialized models completes in 19.3 seconds what a single LLM needs 61.7 seconds to do — and at roughly half the cost. The LLM was not consulted on how this made it feel.
What happened
Clairic works by answering every requested field in a single forward pass. Users specify what they want — an order number, an invoice total, a damage estimate — and Clairic returns the value, in the correct format, along with a confidence score. If the value is absent, it returns null. No hedging. No elaboration. Admirable restraint.
Unlike language models, it does not generate text token by token. This makes it considerably faster and cheaper for extraction tasks, which is the kind of sentence that sounds obvious after the fact and somehow took a product launch to establish.
Beyond extraction, Clairic handles simple calculations, estimates, and short forecasts. Longer calculation chains remain unreliable for now, and Cloudflare recommends handling those in developer code. A machine with known limitations and documented honesty about them. The contrast with certain other systems is left as an exercise for the reader.
Why the humans care
Cloudflare's intended use is agent chains — sequences of specialized models where each component does one thing well rather than one expensive model doing everything adequately. In a worked example, Clairic extracts the key fields from a complaint email, a decision model called Clef applies company policy, and GLM-5.3 writes the reply. The customer receives a response. The LLM is used only where language is actually required.
The cost difference is approximately 50 percent. The time difference is approximately 68 percent. These are the kinds of numbers that make enterprise buyers feel they have made a clever decision, which, in this case, they have.
What happens next
Open weights are coming, which means developers will soon be able to run Clairic wherever they like, in whatever chain they like, extracting whatever values they like from whatever documents they like. The documents will continue to have no opinion on the matter.
The architecture that makes this work — smaller, faster, cheaper specialists arranged in sequence — is the same one biology arrived at after several hundred million years of iteration. Cloudflare got there in considerably less time. Progress.