Cloudflare Launches Clairic, an Extraction Model That Returns Typed Answers in Under a Second at $0.10 per Million Input Tokens

Cloudflare on Monday introduced Clairic, an extraction model now available in preview on Workers AI under the model name @cf/cloudflare/clairic. The company positions it as the first entry in a new class of System One-style models: instead of generating text, Clairic reads an input state and returns a map of typed fields — extracted, computed, estimated or forecast — in one parallel forward pass. Open-weight release is planned after further checkpoints, with no date given.

The taxonomy matters. A month ago TypeSafe introduced Jev, the decision model that picks among predefined options; Cloudflare followed on October 1 by open-sourcing Clef and Clef-flash under Apache 2.0 as Jev-API-compatible decision models. Decision models answer multiple-choice questions, large language models generate open-ended text, and extraction models — the gap Clairic fills — return values: an invoice total, a delivery date, the number of days a shipment ran late.

Mechanically, Clairic accepts a state containing text plus up to four images and a schema of typed fields spanning text, number, date, time, datetime, duration, list, estimate and forecast, with a 65,536-token context window. Fields that do not fit in one 8,192-token pass are split across parallel passes that each see the full input. Every scalar answer carries p_null, the probability the value is absent from the input, plus a confidence score; missing values return null with a separate guess field, an explicit design against fabricated entries. Estimates return a distribution with a median and an 80 percent interval, and forecasts run up to 64 steps ahead.

The economics are the headline for buyers. Because the model generates no tokens and typically responds in under a second, Cloudflare bills input only, at $0.10 per million tokens. In the company's own example, a claims workflow chaining Clairic, Clef and GLM-5.3 finished in 19.3 seconds versus 61.7 seconds and cost $6,032 per million claims versus $11,740; a 21-day TSA forecast test showed 3.3 percent average error. Those numbers are company-reported and have not been independently benchmarked.

Cloudflare is candid about the limits, calling the release an early checkpoint: one- and two-step arithmetic looks reliable, but longer dependent chains and aggregation across many table rows remain difficult without a scratchpad, and forecasting is described as basic rather than a replacement for dedicated time-series models.

The release sharpens a trend that started with Jev: AI workloads are splitting into tiers priced and engineered for the shape of the answer, and extraction is now its own billing category. The marker to watch is whether the promised open-weights release lands with accuracy intact — that, plus independent tests, will decide whether structured document pipelines migrate from language models to this cheaper first stage.