Cloudflare Releases Clairic Extraction Model on Workers AI

Clairic operates by answering every requested field in a single forward pass, without generating text or consuming output tokens. The model supports a 65,536-token context window and can process up to four images per request, running on Cloudflare-hosted GPUs via Workers AI.

In Cloudflare's internal tests, Clairic achieved sub-second extraction times, with the extraction step taking 0.46 seconds compared to 16.1 seconds for GLM-5.3 in their claims pipeline. However, the model showed unreliability with multi-step arithmetic, and 4 out of 21 TSA next-day forecasts fell outside its 80% interval. No standard accuracy benchmark has been published.

Clairic builds on the System One concept from TypeSafe's Jev, which is inspired by Kahneman's fast, intuitive thinking. While Jev and Cloudflare's Clef are decision models that score pre-listed options, Clairic handles answers that cannot be pre-listed, such as invoice totals or repair costs.

A Clairic request includes input text and up to four images, along with a map of named, typed fields. Supported types include text, number, date, time, datetime, duration, list, estimate, and forecast. Each scalar answer includes a probability of absence (p_null) and a confidence score, while text answers provide exact spans with character offsets.