Cloudflare's extraction model Clairic entered preview on Workers AI on October 19, 2026: it generates no text at all. It reads a passage of text or images once and returns typed values directly — amounts, dates, durations, lists, or an estimate or forecast with an interval around it. Cloudflare has given this kind of model a new category name: System One extraction models.
It does a different job from a chatbot
A large share of today's structured-extraction calls are really LLMs being forced into the wrong job: ask a model for an order number, a delivery date, a tax-inclusive total and days overdue from one email, and it has to generate JSON token by token, may miswrite a digit or break the format, and still needs parsing, validation and retries afterwards. Clairic's request shape is entirely different: the input is a state (the body text plus up to four images) and a set of fields with declared types, and one forward pass answers every field in parallel. Text fields come back as source spans with character offsets; numbers and dates carry a confidence score and p_null, the probability that the value is not in the input at all; when a field is absent, the model returns null instead of inventing a plausible answer. In Cloudflare's example, a support email never states the tax-inclusive total, yet Clairic added the $649.00 price and the $52.10 tax into 701.10 within the same pass; a receipt photo with no discount on it correctly returned null for that field.
The time and money saved, in Cloudflare's own ledger
Cloudflare benchmarked a damaged-goods claim pipeline as a cascade: Clairic extracts, Clef decides, GLM-5.3 writes the reply — 19.3 seconds end to end. Giving all three steps to GLM-5.3 took 61.7 seconds, with extraction and decision alone consuming 38 seconds, while Clairic plus Clef finished in under a second combined. Total cost fell from $11,740 to $6,032 per million claims, a saving of 49%. Forecasting was measured too: over the U.S. Transportation Security Administration's daily screening series, Clairic predicted the next day's count for 21 consecutive days with a 3.3% average error, and 17 of the 21 actual counts landed inside its 80% interval; it missed noticeably over the Columbus Day weekend, because a series alone cannot reveal a holiday. The cascade connects neatly with the Cloudflare Clef decision model covered earlier on this site: one model reads the facts accurately, one picks among options, and only then does an LLM write the words.
The limits are printed in the release notes — it is not a universal calculator
Cloudflare lists the weaknesses itself: multi-step arithmetic is unreliable, long chains of dependent calculations and aggregation across many table rows can go wrong, and the official advice is to let code do the math while the model only reads values. Estimates need comparables or history underneath them; without evidence there is no credible interval. Forecasting is basic and is not meant to replace dedicated time-series models. For now Clairic is a hosted preview on Workers AI as cf/cloudflare/clairic, priced on its model page at $0.10 per million input tokens with no output charge, because it produces no output tokens at all; open weights are planned but undated. For teams processing thousands of invoices, tickets and emails a day, its real value is not replacing LLMs — it is carving the most repetitive, format-sensitive layer of calls out of the LLM bill.
