Cloudflare introduced Clairic on 19 October as a preview that extracts values from text and images into specified types, such as numbers and dates. That promises less work turning a model’s answer into usable data. Early tests expose a separate job: deciding whether a value is correct, absent or simply missed. A Polish practitioner’s nutrition-label report and our small check of the public demo make that distinction unusually concrete.
Based on 7 reviewed sources · See sources
A number is usable before it is trustworthy
Clairic takes unstructured input and a map of requested fields. Its typed outputs can give an application a date or amount without asking a text-generating model to write an answer and then parsing it. Cloudflare’s announcement below introduces that narrower approach.
Cloudflare announces its Clairic preview. This is the vendor’s description of the model, not independent evidence of performance.
View the original post on X ↗It differs from the fixed-choice decisions in Microsoft’s Decision-1: the developer specifies the kind of value to extract, rather than supplying every possible answer. Yet a valid date can still be the wrong date. Cloudflare’s extraction guide explicitly separates predictable output structure from factual correctness.
The response also separates a candidate’s confidence from the estimated probability that the requested information is absent, called p_null. The model documentation illustrates the distinction with a missing purchase-order number: a highly confident guess accompanies a null value. Promoting the guess on confidence alone would undo that useful abstention.
The label test: useful confidence, troublesome absence
Warsaw-based developer Tom Smialowski reports testing 495 photos of nutrition labels, requesting nine nutritional values from each and comparing readings with values he had manually checked. He reports 93.2% of printed values read correctly. That is his reported result for printed values, not the percentage of labels read perfectly.
His confidence breakdown is encouraging within that sample: among 1,803 answers scoring at least 0.9, he found nine errors. But seven were zeros returned for fibre that was absent from the label. Across 220 labels without fibre, he reports 88 supplied values where the field should have remained empty. He says the separate presence score often helped distinguish these cases.
He selected those confidence thresholds after the test; they are not a validated acceptance policy for other documents. The inspected post does not supply the full photo corpus or raw results. It also discloses unequal inputs and a reference-set advantage for his own scanner, so its scanner comparison cannot establish a fair model ranking. The original report below gives both the promising readings and those limitations.
Smialowski’s personal nutrition-label test, with stated sample, fields and reference checks. These are author-reported observations, not our replication of his image test.
View the original post on X ↗Our demo check left printed IDs blank
We tried Cloudflare’s public extraction demo with three synthetic messages: an English invoice, a German invoice and a coffee invitation. We requested vendor, invoice number, amount, due date and purchase-order number, then repeated the unchanged batch. Both runs returned the expected vendors, amounts and explicit due dates. Both left the two invoice-number cells empty despite IDs being present.
The purchase-order fields were correctly empty, as were all fields for the non-invoice message. The interface displayed a 95% confidence threshold; its reference date remained 16 October. Our published test record includes the full inputs, schema, expected values and observations.
This checks the hosted interface, not isolated API responses: we did not inspect the backend request, raw model output or model revision. It cannot identify whether the omissions arose in the model or the application, and three messages cannot establish an error rate. It does show why an empty cell deserves inspection alongside a filled one.
Two unchanged runs, the same visible result
Marketer Fizz’s small text-only public-demo check, 20 October 2026
| Synthetic input | Printed invoice ID → displayed cell | Amount and due date displayed |
|---|---|---|
| English invoice | SYN-100 → empty | 120.5; 2026-10-30 |
| German invoice | TEST-200 → empty | 2,450; 2026-11-05 |
| Coffee invitation | No ID → empty | Both empty |
Preview access, with a verification step
Clairic is available through Workers AI as @cf/cloudflare/clairic. The listed price is $0.10 per million input tokens, with no output-token charge. Cloudflare says it intends to release weights; the announcement is not evidence that downloadable weights are already available.
For a team turning product sheets or incoming documents into records, the practical lesson is to evaluate false additions and false omissions separately on its own material. Preserve the source, keep absence information alongside confidence and check the fields whose errors matter. Our German example tests one number and date format; it establishes no wider European-language accuracy.
Extraction and forecasting need different treatment too. Cloudflare’s guide says estimate and forecast fields can return predictions with little supporting input. Those values should enter a workflow as predictions, never silently acquire the authority of facts read from a document.
Sources
- Introducing Clairic
Cloudflare · Announcement ·
- Clairic model documentation and price
Cloudflare · Documentation
- Extraction models: types, confidence and missing values
Cloudflare · Documentation
- Original Clairic announcement
Cloudflare on X · Announcement ·
- 495 nutrition-label photos: practitioner test report
Tom Smialowski on X · Research ·
- Public extraction demo
Cloudflare · Documentation
- Synthetic demo inputs, settings and two observed runs
Marketer Fizz · Research ·
Written with AI assistance from the sources above. Analysis and practical implications are our interpretation. Our editorial approach.
