Cloudflare previews Clairic extraction model
Cloudflare has previewed Clairic, a single-pass "System One" model that returns typed values from unstructured input, pitched as faster and cheaper than routing repetitive structured-output calls to frontier LLMs.
Cloudflare announced a preview of Clairic, which it calls the first of a new class of "System One" extraction models. It is hosted on Workers AI as @cf/cloudflare/clairic, and the company says it plans to release open weights soon.
The label comes from TypeSafe, which introduced System One models a month ago with Jev. Jev and Clef are decision models: they return probabilities over a fixed list of options. Clairic covers a different job. It pulls values out of the input, computes them from it, or predicts values the input doesn't state.
According to Cloudflare, Clairic can return four kinds of output:
- Extracted text spans, such as names or order IDs, with character offsets and confidence scores
- Computed numbers, dates, durations and lists that follow type rules, such as two decimals for cents
- Estimated values, returned as distributions within declared bounds, with a median and an 80% interval
- Forecasts of a series, one estimate per step with its own 80% interval
Clairic generates no tokens. It processes the input in a single forward pass, so Cloudflare charges only for input. The company says responses typically arrive in under a second, while an LLM reasoning through the same step takes several seconds. Every answer is a value of the declared type, or null if the model can't find one. Cloudflare says this rules out mistyped outputs, though the model can still be wrong.
A request contains text plus up to four optional images, along with a map of named, typed fields. Cloudflare says Clairic is more accurate than LLMs for certain use cases but gave no benchmark figures or pricing in the material reviewed.