Cloudflare has three AIs work in relay to handle claims: self-tested 3.2 times faster, costs reduced by 49%

Beating AI News Flash: Cloudflare releases preview of data extraction model Clairic. Give it an invoice or an after-sales email, and it can identify order numbers, dates, and total amounts, as well as estimate repair costs and predict customer traffic based on existing information.

It does not need to generate answers token by token like a regular LLM, but instead directly calculates and returns results of the specified type, along with confidence scores. Fields with no supporting evidence can return null.

Clairic continues the System One model approach that emerged after TypeSafe launched Jev. Cloudflare already launched the decision model Clef on October 1, which is responsible for choosing from preset options; Clairic is responsible for extracting, calculating, and predicting specific numerical values. Neither needs to generate answers word by word. Clairic charges only $0.10 per million input tokens, with no output token fees.

Cloudflare demonstrated this with merchandise damage claims: Clairic first extracts information and estimates repair costs, Clef decides whether to refund or repair, and GLM-5.3 finally writes back to the customer. The three working together took 19.3 seconds, while handing everything to GLM-5.3 alone required 61.7 seconds. Calculated per million claims, the cost drops from $11,740 to $6,032, a reduction of 49%. Looking only at the information extraction step, the time drops from 16.1 seconds to 0.46 seconds.

However, although Clairic does not generate output tokens, when fields cannot be processed in a single calculation, the system automatically splits them into multiple parallel rounds, with each round re-reading the full input. For high-token inputs such as images, this may increase actual invocation costs.

The model is currently available in preview on Workers AI, and the weights have not yet been open-sourced.