Introducing Clairic, a new System One class of extraction models

Cloudflare AI has opened a preview of Clairic, the first model in what it calls a new class of System One extraction models. It is available hosted on Workers AI under the identifier @cf/cloudflare/clairic, and Cloudflare says it intends to release the model open-weights soon. The pitch is that Clairic extracts, computes, estimates and forecasts values from unstructured inputs with higher accuracy, lower latency and lower cost than a conventional LLM.
The framing follows TypeSafe's System One announcement about a month earlier, which introduced Jev and drew on Daniel Kahneman's fast-versus-slow distinction from Thinking, Fast and Slow. Until now every System One model in this lineage was a decision model: given an input state plus a list of candidate questions or options, it returns probabilities. Cloudflare's argument is that many structured-output calls to frontier LLMs do not fit that shape at all — invoice totals, delivery dates, days late, repair costs, next-day traffic — because their answers cannot be enumerated in advance, so a decision model has nothing to assign probabilities to.
Clairic keeps the single-pass, typed, input-only billing model but returns values instead of choices: text spans copied from the input with character offsets, computed numbers, dates, times, datetimes, durations and lists, estimated values as distributions inside caller-declared bounds, and forecasts of future steps in a series. Cloudflare positions the three families side by side: Clairic predicts values (regression and extraction), Clef and Jev make decisions among listed options (classification), and LLMs write and reason (generation). Because Clairic produces no output tokens, it cannot mistype a number while generating it or hallucinate in the LLM sense, though Cloudflare acknowledges it can still make mistakes.
A request consists of a state — the text to read plus up to four optional images — and a map of named fields, each with a type and optional plain-language instructions. Supported types are text (a span with offsets), number (honoring integer and decimals), date, time and datetime (ISO form), duration (in a chosen unit), list (matching text, number or date in input order), estimate (a distribution inside [min, max] on a linear or log step grid) and forecast (one estimate per future step). Every scalar answer carries p_null, the probability that the value is not present in the inputs, plus a confidence score. Estimates return a median and an 80% interval from p10 to p90, with all 99 percentiles available on request. Fields are not capped: if they do not fit alongside the input in one pass, the server splits them into parallel passes that each see the whole input and merges the answers. Cloudflare says asking for thirty fields costs roughly the same as asking for three. Outputs are always a value of the declared type or null, and null responses include the model's fallback guess.
The worked examples show the behavior. From a customer support email, a request for six fields returns the order ID as a span with offsets 10 to 17, the delivery date as 2026-09-28, and an amount paid of 701.1 computed by adding the stated $649.00 and $52.10 tax in the same pass. A duration field derives 11 days between delivery and the email date, and a list field returns the two damage descriptions as exact spans. A tracking number that is absent comes back null with p_null of 1 rather than a plausible fabrication. On a photographed receipt, Clairic pulled order number, datetime, quantity, unit price, subtotal, sales tax and tip with confidence scores ranging from about 0.78 to nearly 0.99999, and returned null for a discount that was not shown.
For estimation, Clairic takes the specified range and grid and returns a distribution rather than a single figure. In a comparable-sales example asking for a sale price bounded between $50,000 and $5,000,000, it returned 418,000 with a median of 415,000 and an 80% interval of $370,000 to $470,000. Cloudflare contrasts this with asking an LLM for a number, which yields no sense of certainty, and with approximating an estimate by bucketing options in a decision model, which is coarse and adds tokens per bucket. The width of the interval is itself useful signal for routing a second check.
Cloudflare tested forecasting against the public TSA checkpoint traveler series, asking each morning for the next day's count over 21 consecutive days. On one day it predicted a median of 2,419,283 against an actual 2,419,474. Across the run, forecasts were off by 3.3% on average and 17 of 21 actuals landed inside the 80% interval. The four misses included three days on the Columbus Day weekend, which Cloudflare attributes to holiday travel the series alone does not reveal. Cloudflare is explicit that Clairic does not replace a dedicated time-series model; it sits between per-metric pipelines and pasting history into an LLM.
In practice, the design shifts where errors surface. Because output is typed and bounded, schema validation and retry loops shrink, and offsets make every extracted string traceable to its source. The cost model is the bigger structural change: per-document pricing becomes a function of input length rather than field count, which favors wide schemas over narrow ones. The trade-offs are real. Confidence scores and p_null are probabilistic, so teams still need thresholds and escalation paths for low-confidence results, and Clairic's claimed accuracy advantage is measured on the tasks Cloudflare chose to show. Forecasts inherit whatever patterns the series contains, as the holiday misses illustrate.
Why it matters: Teams currently paying frontier-LLM prices to pull numbers, dates and line items from documents, emails and images have a cheaper, lower-latency option that returns typed values with confidence and traceable spans. Expect the architectural pattern to become a pipeline — Clairic to extract, a decision model to classify, an LLM to explain — with routing rules keyed on confidence intervals. That is inference from the stated design, but it follows from billing input tokens only and returning null instead of guessing.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Clairic is a new model from Cloudflare that is a System One extraction modeI, capable of extracting, computing, estimating, and forecasting data from unstructured inputs.