Cloudflare has added a new model to Workers AI: Clairic, published under the identifier @cf/cloudflare/clairic and described in the company's changelog as the first "extraction model" on the platform. The entry is dated 2026-10-19. Clairic is available as a preview, and Cloudflare states that it intends to release the model with open weights soon, though no date is given.
Choosing versus extracting is the distinction Cloudflare draws. The changelog explains that decision models such as Clef pick among options defined up front. Structured-output tasks whose answers cannot be listed in advance are what Clairic targets — the source names a delivery date, an invoice total, the cost of a repair, or the number of days a shipment is late. In those cases the caller supplies a state plus a schema of typed fields, and one typed value per field is returned by the model.
Mechanically, Clairic does not generate tokens. According to the changelog, it reads the input in a single pass and decodes each answer under the rules of its declared type, so each result is either a valid value of that type or null when the input does not determine it. Cloudflare categorises the returned values three ways: read from the input, computed from it, or predicted beyond it.
How many fields a caller defines has no upper limit. The changelog states that any fields which cannot be placed alongside the input in a single pass get divided among parallel passes, so a single context window does not cap field count the way one might expect. Up to four images may also be supplied in addition to the state.
Each scalar answer arrives with two accompanying numbers: a confidence score, and p_null — the likelihood that the value is not determined by the input. For estimates and forecasts, a median plus an 80% interval running from p10 to p90 are also returned, and setting options.include_quantiles brings back all 99 percentiles. Cloudflare describes these as the foundation for choosing what to automate and what to send to human review.
The changelog's own example is an invoice: a state string describing invoice INV-2291 from Northwind Supply, issued 2026-09-02 on net 30 terms, with four pallets at $250.00 each and $80.00 shipping, plus two contact names. The schema asks for the invoice number and issue date as text and date, a due date with the instruction "Issue date plus the net payment days," a two-decimal USD total including shipping, and a list of contacts. The documented response returns "INV-2291" for the invoice number, with its span in the state, and the amount due for the total, alongside p_null and confidence.
Where the preview checkpoint falls short is stated explicitly by Cloudflare, and those caveats matter more than the demo. Totals, differences, conversions and date arithmetic — one- and two-step calculations — are described as reliable, but aggregations across many table rows and long chains of dependent steps are not yet. Extracting the inputs with Clairic and doing the math in code is the recommended workaround.
Estimates are subject to bounds in a similar way. Clairic estimates best, per the changelog, when the input carries evidence like comparables, history or descriptive text; absent such evidence, the interval comes out wide. Forecasting is described as basic: next-step forecasts follow recent patterns, yet forecasts several steps out flatten toward the average, and the model does not anticipate events the series does not show, holidays among them. Cloudflare states plainly that it does not replace a time-series model trained on your own data.
Correctness and calibration get separate treatment. Intervals are trained to be calibrated and are described as being so on most held-out sets, yet on unfamiliar data they can run narrow. As for confidence, the changelog warns that every answer is well formed, not necessarily correct — a sentence that should shape how anyone wires the model into an automated pipeline.
Dates carry their own rule. Dates without a year are resolved from context, so a field that depends on the current date requires either a reference_date parameter or a line such as "Today: 2026-10-19" added to the state. That is a small implementation detail with outsized consequences for anything computing deadlines or lateness.
Pricing is stated per input token only: $0.10 per million input tokens, with no output token charge because the model generates none. Access is through the Workers AI binding via env.AI.run() or the REST API at /ai/run, and Cloudflare says AI Gateway can be used with these endpoints.
For developers building document and data pipelines, the practical shape of the offer is a schema-first extraction step that returns machine-checkable types rather than free text to be parsed. The p_null and confidence fields, plus quantile output for numeric estimates, give a pipeline somewhere to branch: auto-accept high-confidence values, queue the rest. That is editorial interpretation of the documented fields, not a claim Cloudflare makes about accuracy in production.
The tradeoff is that the reliability envelope is narrow and stated. Multi-step arithmetic and row-level aggregation should stay in code, per the source, and forecasting should not displace a purpose-trained series model. Anyone treating Clairic as a general reasoning endpoint would be working against the documented limits.
What remains unknown is substantial. The open-weights release is announced as an intention with no date. The changelog calls the current endpoint an early checkpoint and says Cloudflare expects it to improve, without committing to a schedule or to which weaknesses will close first. No independent evaluation of the model's accuracy is provided in the source, and the calibration claim is Cloudflare's own description of held-out performance.
The sensible reading for a working developer is to treat Clairic as a preview extraction primitive with unusually candid documentation. Its value depends on whether the fields you need are read or computed in one or two steps from text you already have, and whether your pipeline can act on p_null and confidence rather than assuming every returned value is right.