Cloudflare's Clairic extracts values instead of writing text

Cloudflare announced Clairic on Monday, October 19, describing it as its first System One extraction model. Its announcement cites TypeSafe's Jev and Cloudflare's Clef as decision models that choose among supplied options. Clairic instead returns typed values such as numbers, dates, text spans, lists, and distributions. The category framing and first-model description are Cloudflare's.

Clairic is available as a hosted preview on Workers AI under model id `@cf/cloudflare/clairic`, through a Worker binding or the REST API. Cloudflare says it intends to release open weights soon but gives no date; this review verified hosted access documentation, not an available weights release.

A request supplies text, up to four optional images, and named fields with types and instructions. Supported types include text, number, date, time, datetime, duration, list, estimate, and forecast. Cloudflare describes parallel decoding of typed values rather than output-token generation. Its model page lists a 65,536-token context window, but also says fields that do not fit beside the state in one 8,192-token pass are split across parallel passes. A request is therefore not guaranteed to use only one pass.

Fields extracting or computing a single value include `p_null` and a confidence score. A null means the model considers the input insufficient; it can also mean the model failed to identify an answer that was present. In Cloudflare's missing-tracking-number example, the returned value is null while a separate guess is retained. Estimates and forecasts return distributions, including a median and p10–p90 interval. None of these signals guarantees correctness. Cloudflare lists $0.10 per million input tokens with no output-token charge; its typical sub-second latency is a vendor claim, not a service guarantee or a measurement reproduced here.

Cloudflare compared a demonstration claims pipeline—Clairic extracting values, Clef deciding, and GLM-5.3 writing—with an all-GLM-5.3 version. It reports 19.3 seconds versus 61.7 seconds end to end, about 69% less time, and $6,032 versus $11,740 per million claims, about 49% lower cost. Those costs extrapolate the example workload; they are not guaranteed savings for every claims system. Cloudflare reports under a second combined for extraction and decision, versus 38 seconds in its all-LLM comparison. Its stable repair-cost estimate across repeated runs demonstrates consistency in that example, not proof that the estimate matches an actual repair bill. These are vendor-reported results, not tests reproduced for this briefing.

For forecasting, Cloudflare reports a 21-day test using public TSA checkpoint screening counts, with average error of 3.3% and 17 actual counts inside the model's predicted 80% intervals. It says three of the four misses occurred over Columbus Day weekend. This small vendor-reported test does not establish general forecast accuracy or calibration, and its predicted intervals are not guarantees about future outcomes.

Cloudflare calls the endpoint an early checkpoint. It says short arithmetic looks reliable in its tests, while longer dependent calculations and aggregation across many rows remain limitations; it recommends extracting the inputs and doing the math in code. Estimates also need supporting evidence such as comparables or history. These are the vendor's stated capabilities and limits, not independently validated reliability findings.

Cloudflare says intervals are trained for calibration but may be too narrow on unfamiliar data. Its basic forecasting follows recent patterns and can flatten toward the average several steps out; it is not a replacement for a time-series model trained on the user's data. A valid output type is separate from a correct value. Confidence, null probabilities, and intervals can help route review, but important calculations and decisions still need workload-specific validation.

Cloudflare's example separates value extraction, decision selection, and reply writing. The final LLM sees selected facts rather than the raw email, which can reduce its direct exposure to untrusted instructions. That does not establish prompt-injection immunity or remove the need to validate extracted values, decisions, and generated replies. The announcement links a live demonstration of this pipeline.

Our view: typed extraction is worth evaluating where applications currently ask an LLM for structured data. Constraining output types can simplify parsing and format checks, but it does not eliminate misread values, incorrect computations, or unsupported estimates. The useful test is accuracy, latency, and completed-task cost on representative documents. Clairic remains in preview; the reported pipeline and forecasting results are examples to investigate, not promises.