Cloudflare Previews Clairic, a System One Extraction Model on Workers AI

A New Class of System One Models

Cloudflare has announced a preview of Clairic, the first model in a new class of System One-style extraction models. The model is available hosted on Workers AI as @cf/cloudflare/clairic, and Cloudflare says it intends to release the model open-weights soon. The announcement follows TypeSafe's introduction of the System One decision model Jev a month ago, a concept drawn from Daniel Kahneman's book Thinking, Fast and Slow, where System One thinking is fast and precise while System Two covers deliberate reasoning like LLMs.

Extraction, Computation, Estimation, and Forecasting

Unlike decision models such as Jev and Clef, which return probabilities against listed options, Clairic answers with values: extracted, computed, estimated, or forecasted. Extracted values come back as spans with character offsets and confidence scores. Computed values such as numbers, dates, durations and lists follow the rules of their type. Estimated values return as a distribution inside user-declared bounds, with a median and an 80 percent interval. Forecasted values give one estimate per step ahead, each with its own interval. Because the model generates no tokens, Cloudflare says it cannot produce mistyped answers or hallucinate in the way an LLM can, though it can still make mistakes. Outputs are always a value of the declared type, or null when the model fails to find an answer.

One Pass, Many Fields

A request consists of a state (the text to read, plus up to four optional images) and a map of typed fields, each with a type and optional plain-language instructions. Every field is answered in a single forward pass with no reasoning and no output tokens, so the model charges only for input and typically responds in under a second. There is no limit on the number of fields; if they do not fit beside the input in one pass, the server splits them into parallel passes and merges the answers. Cloudflare demonstrated Clairic on a customer support email, a photographed receipt, a home price estimate based on comparable sales, and a forecast of daily TSA checkpoint screening counts. Over 21 days of TSA forecasts, the model was off by 3.3 percent on average and 17 of the 21 actual counts landed inside its 80 percent interval.

A Pipeline with Decision Models and LLMs

Cloudflare positioned extraction models as one step alongside decision models and LLMs: Clairic turns raw input into typed values, Clef makes decisions from those values, and an LLM writes the response. In a live demo pipeline for a damaged-product claim, the cascade finished in 19.3 seconds against 61.7 seconds for an all-LLM version, about 3.2 times faster, at a cost of $6,032 per million claims versus $11,740. Each step also receives cleaner, typed input than the one before, and the LLM only sees the decision and the facts it needs, not the raw email.

Sources: Cloudflare