Cloudflare previews Clairic, a typed-value extraction model on Workers AI

Cloudflare's Clairic, in preview on Workers AI, returns typed fields with confidence scores from text and images instead of generating text.

Cloudflare has put Clairic, which it calls a System One extraction model, into preview on Workers AI as @cf/cloudflare/clairic, according to its October 19 announcement. You send text, up to four images and a map of named, typed fields such as a total, a date or a list. You get back one value per field in a single pass. Each value is the type you asked for, or null, with a confidence score and p_null. If you currently prompt an LLM to return invoice or receipt fields as JSON and then validate the output, this targets that step.

In Cloudflare's examples, Clairic added a price and tax into a total the email never stated. It returned null for a missing tracking number. It read a skewed receipt photo, giving lower confidence on the order number and purchase time than on the amounts, as DEV Community's summary notes. The Workers AI changelog lists the price as $0.10 per million input tokens, with no output charge. You can call it through env.AI.run() or the REST API.

The caveats come from Cloudflare itself. The changelog calls the endpoint an early checkpoint. One- and two-step arithmetic is reliable, but long dependent chains and sums across many table rows are not, so Cloudflare advises doing that math in code. It says every answer is "well formed, not necessarily correct." PPC Land points out that the accuracy claims rest on Cloudflare's own examples and tests, not a general benchmark. Open weights are promised "soon" with no date. Treat the confidence scores as a routing signal for human review and test them on your own messy documents.

Sources