Cloudflare: clairic

Clairic is an extraction model from Cloudflare, hosted on Workers AI under the model ID @cf/cloudflare/clairic. It takes text, optional images, and a typed field schema as input and returns strongly typed values in a single pass — text spans, numbers, dates, times, durations, lists, estimates, and predictions — along with a confidence score and null probability for each field. It does not generate output tokens. Cloudflare positions it as a new System One class of AI model and as the first model applying that concept to extraction tasks, inspired by the Jev model announced by TypeSafe. The model is built for pulling structured data out of documents such as invoices, receipts, and insurance paperwork. It also supports financial data aggregation, date and duration calculation, inventory forecasting, and information extraction in customer service. Cloudflare reports higher accuracy, lower latency, and lower cost than conventional LLMs on data extraction tasks, and says it handles one-step or two-step calculations such as sums, differences, conversions, and date math reliably. In a three-stage pipeline test for insurance claim processing, it reduced workflow time from 61.7 seconds to 19.3 seconds, a 69% latency reduction, while cutting cost by about 49%; it also succeeded at estimating repair costs where Gemma 4 and Qwen pipelines failed. Clairic is intended as a component of a broader AI architecture rather than an LLM replacement, and works alongside LLMs and the Clef decision model in a three-stage pipeline. It is best suited to well-defined extraction and simple calculation tasks, and should be combined with LLMs for complex reasoning and text generation. The preview version still struggles with multi-step arithmetic and long dependency chains or aggregation across multiple table rows. Estimates depend on comparable data, history, or descriptive text in the input and are only limitedly accurate without evidence, while prediction is basic and does not replace time-series models trained on specific data. It is meant for general pattern matching and non-critical uses. Clairic is currently available as a preview, with an open-weights release planned in the near future.