Cloudflare has launched a preview of Clairic on Workers AI, giving developers a model that extracts information, calculates results and predicts values from unstructured material. Requests incur charges for input tokens only, and Cloudflare says System One models typically answer in less than a second. Developers can access Clairic as @cf/cloudflare/clairic.
Clairic accepts text alongside as many as four optional images, with developers specifying the fields they want answered and the type of each result. A request can also include instructions written in ordinary language for individual fields. Cloudflare plans to make the model's weights available soon, making the hosted preview the current route to access.
Read more: Cloudflare Patches Containers Flaw That Could Expose Disk Data
The model targets jobs such as finding an invoice total, calculating how late a delivery was or estimating a repair bill. Cloudflare describes Clairic as an extraction member of the System One category, processing requests in a single pass without generating answer tokens or performing reasoning. That gives it a different role from Clef, which selects among supplied options and assigns probabilities to them.
For text extraction, Clairic returns a passage taken from the input together with character positions that identify its location and a confidence score. It can also return calculations and calendar values in the requested data type, including rules such as decimal precision for amounts. Cloudflare says answers follow the declared type or return null when the model cannot locate an answer, although the model can still make mistakes.
For estimates, developers set numerical limits and the spacing of possible values, and Clairic returns a distribution with a median and an interval covering 80%. Forecasting produces a separate estimate and interval for each future step in a series. Developers can also request all 99 percentiles for an estimate to inspect its distribution in more detail.
Cloudflare places no cap on how many fields developers can request. When the fields cannot fit alongside the input in one pass, the server divides them across parallel passes, gives each the complete input and combines their results. Scalar results include confidence information and a probability that the requested value is absent from the supplied material, giving developers signals for deciding which answers need review.
