The Pulse
Cloudflare Previews Clairic, a Typed Data Extraction Model
Cloudflare introduced Clairic on October 19 as a preview model on Workers AI for extracting, calculating, estimating and forecasting typed values from unstructured input. The company says it runs in a single pass, but warns that its early c
AI.info Team · Written with AI · fact-checked ·
Cloudflare introduced Clairic on October 19 as a preview model on Workers AI, built to pull typed data from unstructured text and images without generating a written response. The company positions it as a new kind of “System One” model: a fast, single-pass tool for finding or calculating values that might otherwise be sent to a general-purpose language model. Cloudflare says the hosted endpoint is available as @cf/cloudflare/clairic; it plans to release open weights later.
Clairic targets jobs such as identifying an invoice total, calculating how many days have passed since delivery, estimating a repair cost or forecasting a near-term value from a time series. Rather than return free-form text, it responds in fields developers define, with types such as text, number, date, duration, list, estimate and forecast. Its answers can include confidence scores, a probability that a requested value is absent, and, for estimates and forecasts, an 80% interval.
One pass, with typed answers
A request gives Clairic a block of text, optionally accompanied by up to four images, and a set of named fields with types and instructions. Extracted text can include character offsets pointing to where a value appeared in the input. Missing answers can return as null, while numerical fields can follow rules such as a fixed number of decimal places.
Cloudflare says the model processes the input in one forward pass and charges only for input tokens, not generated output tokens. Requests with many fields may be divided into parallel passes if the fields do not fit alongside the input in one pass. The company says those passes each see the full input before their answers are merged.
“The current endpoint is an early checkpoint and we expect it to get better,” write Alex Reneau, Michelle Chen and Kevin Flansburg, the authors of Cloudflare’s announcement.
Cloudflare’s tests favor a model cascade
To show how Clairic could fit into a customer-service workflow, Cloudflare compared a three-step sequence—Clairic for extraction, its Clef model for a decision, and GLM-5.3 for a written reply—with a version that used GLM-5.3 for all three steps. In the company’s test, the sequence finished in 19.3 seconds, compared with 61.7 seconds for the all-GLM workflow. Cloudflare calculated a 49% lower cost per million claims for the cascade: $6,032 versus $11,740.
Those results describe Cloudflare’s specific example, not a general benchmark across tasks or providers. The company says both workflows using GLM-5.3 got the facts and decision right in every run, but the repair estimate from GLM-5.3 varied between runs while Clairic’s estimate did not. Cloudflare also reports that Gemma 4 and Qwen failed to infer some values in the same task.
Cloudflare separately tested Clairic’s forecasting on daily U.S. airport checkpoint counts from the Transportation Security Administration. Across 21 one-day-ahead forecasts, it reports an average error of 3.3%, with 17 actual counts falling within the model’s 80% intervals. The four misses included three days during the Columbus Day weekend; Cloudflare says the model could not account for holidays that were not reflected in the historical series.
Cloudflare warns against trusting every well-formed answer
The company describes Clairic as an early checkpoint and spells out limits that matter for deployment. One- and two-step calculations appear reliable in its initial testing, but longer chains of dependent arithmetic and aggregation across many table rows can be difficult. Cloudflare recommends using ordinary code for the arithmetic after Clairic predicts a value.
Estimates depend on evidence in the input, such as comparable examples or historical data. Cloudflare also cautions that confidence intervals can be too narrow on unfamiliar data and says the model is not intended to replace specialized time-series forecasting systems for detailed, long-horizon work. A typed result constrains the shape of an answer; it does not guarantee that the answer is right.
Developers can call the preview from a Worker or through Cloudflare’s REST API. For now, Clairic’s clearest role is as one step in a larger workflow: extract and structure values, pass those values to a decision model or application code, and reserve a language model for the response that actually needs to be written.