Cloudflare previews Clairic, a System One model that extracts and estimates typed values

Clairic is the first model in a new class Cloudflare calls System One extraction models, which sit alongside its existing decision models Jev and Clef. Each request sends text, up to four optional images and a map of typed fields, and returns one answer per field from a single parallel pass. Scalar answers include a probability that the value is absent from the input, along with a confidence score.

Estimate fields return a median and an 80% interval inside declared bounds, and forecast fields return a value and interval for each future step. Cloudflare's examples include pulling an order number and damage list from a support email, estimating a sale price from comparable sales, and forecasting daily TSA checkpoint counts. In the TSA test, forecasts were off by 3.3% on average over 21 days, and 17 of 21 actual counts fell inside the 80% interval.

Cloudflare also lists limits for the early checkpoint. Multi-step arithmetic and aggregation across many rows are not yet reliable, so it suggests having the model predict values and letting code do the math. Estimates work best when the input contains comparables, history or descriptive text, and forecasting is basic and not a replacement for dedicated time-series models.

Developers can replace some LLM extraction calls with a faster, cheaper single pass that returns typed values, spans and null for missing fields, which code can validate. In Cloudflare's test pipeline, a Clairic and Clef front end finished in 19.3 seconds against 61.7 seconds for an all-LLM version. Treat forecasts and multi-step math as preview-quality, and add checks for any value the input does not clearly support.