Cloudflare released a preview of the Clairic extraction model on Workers AI under the model identifier @cf/cloudflare/clairic.
Cloudflare plans to release open weights for the Clairic extraction model following its initial Workers AI preview.
Cloudflare tests demonstrated that Clairic completed data extraction in 0.46 seconds, compared to 16.1 seconds for GLM-5.3.
Cloudflare priced the Clairic extraction model at $0.10 per million input tokens on Workers AI with no charge for output tokens.
A claims processing pipeline combining Clairic, Clef, and GLM-5.3 completed in 19.3 seconds, compared to 61.7 seconds when GLM-5.3 performed all steps alone.
The Clairic extraction model processes raw inputs and returns typed values in a single forward pass without generating output tokens or text.
Using Clairic and Clef alongside GLM-5.3 reduced estimated costs for a damaged-product claims pipeline by 49%, from $11,740 to $6,032 per million claims.
Cloudflare structured an AI stack where extraction models turn raw input into typed values, decision models determine choices, and LLMs write text output.
Cloudflare framed Clairic as a System One model based on Daniel Kahneman's fast, intuitive thinking concept, expanding System One functionality to field extraction.
Cloudflare reported that Clairic cannot reliably handle multi-step arithmetic, long dependent calculations, or aggregation across many table rows.
In Cloudflare's damaged-product claims pipeline example, Clairic extracted customer details, Clef determined whether to refund or repair, and GLM-5.3 authored the reply.
The Clairic extraction model accepts an input state consisting of raw text and up to four images within a 65,536-token context window.
Clairic supports field types including text, number, date, time, datetime, duration, list, estimate, and forecast.
In a 21-day TSA passenger volume forecasting test, 4 of Clairic's 21 next-day forecasts fell outside its 80% interval, with 3 misses on Columbus Day weekend.
Every scalar answer generated by Clairic includes a confidence score and p_null, which indicates the probability that the field value is absent.
For text extraction fields, Clairic provides exact text spans along with character offsets corresponding to the input text.
Fields exceeding Clairic's 8,192-token single-pass limit split into parallel passes, which can cause schema token costs to exceed the document cost.