CLOUDFLARE has announced a new System One extraction model, Clairic, described as the first in this class. Available in preview on Workers AI as @cf/cloudflare/clairic, Clairic is designed to output typed values directly from unstructured inputs, rather than generating free text or tokens. The company emphasises speed and cost advantages, with single-pass processing that returns extracted values, computed values, estimates and forecasts, each accompanied by appropriate confidence indicators.
The article explains how Clairic sits alongside existing System One models (extraction, decision) and System Two models (LLMs). In practice, Clairic processes a state (text and up to four images) and a map of typed fields, delivering one answer per field in parallel, with spans, p_null, confidence scores and typed output. Examples include extracting an order number, delivery date, amounts, and damage descriptions from an email, or pulling values from a receipt image.
Clairic can also produce estimates and forecasts with error bounds, and a demonstration pipeline shows Clairic handing extraction and decision steps before an LLM generates a reply, delivering notable gains in end-to-end speed and cost compared with an all-LLM approach. The current preview notes constraints on multi-step arithmetic, the need for evidence for estimates, and calibration caveats; open weights are planned for later.