The News, Explained
Cloudflare announced a preview of Clairic, a model for extracting usable values from documents or images, on October 19, 2026. Developers specify the fields they need and types such as numbers or dates, then receive corresponding values. The model is hosted on Workers AI; releasing its weights remains a future plan. Source
Different questions require different evidence
A request combines state, the material to read, with fields defining the requested values. Consider a hypothetical invoice workflow. Reading the printed total is extraction; working out a deadline from an issue date and payment terms is computation. Estimating when payment will actually arrive concerns an unknown outcome. Asking for an outstanding balance at each future weekly step is forecasting. The official documentation distinguishes these four tasks. Source
The model returns values in requested types, including numbers, dates, text and lists, rather than composing an open-ended response. Extracted text includes start and end positions identifying its location in the input. Those source spans let an application show a returned name or order number in its original context for comparison. Source Source
For fields that extract or compute a single value, null means the model considers the input insufficient to determine an answer. Information may be absent, or the model may have failed to find it. Lists are empty when no matching values are found. Estimates and forecasts can still return predictions with little evidence, making the kind of result an essential part of interpreting it. Source
Reading confidence and prediction intervals
Single-value extraction and computation results include confidence scores. Estimates and forecasts include distributions over possible values. The range from p10 to p90 represents the middle 80% of the model’s predicted distribution. Actual outcomes can fall outside it, and its realized coverage can vary across tasks and datasets. Source
Constraints limit the form a returned result can take. In the documentation’s example, restricting a payment-date estimate to November produces a date within November, while the invoice could actually be paid in December. Checking that an answer follows its format and bounds and checking that it is correct are separate tasks. Source
The preview also has arithmetic limits. Cloudflare identifies chains of calculations and aggregation across many table rows as weaknesses, and says intervals can be too narrow on unfamiliar data. It recommends leaving the arithmetic to code and describes forecasting as suitable for basic, general pattern matching. Source
The model reference lists a price of $0.10 per million input tokens. Tokens are the units used to count model input; Clairic generates no output tokens. Its documented context window is 65,536 tokens. Field definitions that do not fit beside the input in an 8,192-token pass are split across parallel passes. Cost assessment therefore needs the actual request’s input usage. Source Source
OYOPICK’s Take
The useful distinction is the evidence behind numbers that may otherwise look identical on a screen. A printed invoice amount can be compared with its source, while an estimated payment date needs evaluation against outcomes collected later. Putting both in the same “confirmed information” field can obscure that difference.
Connecting extracted values to source spans, calculations to arithmetic checks, and estimates to prediction ranges and actual outcomes could make better use of this model’s design. Its disclosed difficulty with longer calculations gives a concrete reason to separate extraction from subsequent arithmetic. If workflow interfaces and records preserve those distinctions, AI could take on more work while giving people clearer places to identify and correct errors.