Cloudflare launches Clairic to extract figures from invoices and email

Cloudflare has released a preview of Clairic, a model that pulls numbers, dates, names and lists out of invoices, receipts, emails and other unstructured documents. The model does not write text the way an ordinary language model does. It answers with typed values, a confidence score and a pointer to where in the document the answer appears. For organisations that pay large language models to read documents today, that could mean lower cost and simpler control. Cloudflare itself calls the model an early version.

Cloudflare published the announcement on its own blog on 19 October 2026 at 19:48 Oslo time. Clairic is available as a preview in the Workers AI service under the name @cf/cloudflare/clairic. The company writes that it intends to release the model with open weights soon, but gives no date.

A new type of model

Cloudflare places Clairic in what it calls System One models. The term builds on Daniel Kahneman's distinction between fast and slow thinking. According to Cloudflare, the term was first used by TypeSafe for its Jev model. Cloudflare already offers the decision model Clef, which chooses between fixed options and gives a probability for each choice.

Clairic is the first extraction model in this family. The user sends in the text, optionally with up to four images, and a list of fields to fill in. Each field has a type: text, number, date, time, duration, list, estimate or forecast. The model answers all fields in one parallel pass, without reasoning and without generating tokens.

Cloudflare states that the model charges only for input and typically responds in under a second. The company also writes that thirty fields cost about the same as three, because everything is computed at once. A concrete per-token price for Clairic is not stated in the announcement.

Answers that can be checked

The most important point for leaders is not speed but traceability. Extracted text comes back with its character position in the source document. Every answer carries a confidence score and a probability that the value is missing. If the value is not in the document, the model returns null instead of a plausible-looking guess.

Estimates come as an interval holding 80 percent of the probability, not as a single number. The width shows how certain the model is. Cloudflare describes this as a basis for deciding what can be automated and what should go to manual review.

Cloudflare shows an example in which a customer complains about a damaged product by email. Clairic extracts the order number, delivery date, amount paid and type of damage, and works out the number of days since delivery. The decision model Clef then chooses between a refund and a repair, and a language model writes the reply. According to the company, the language model never sees the email itself, which Cloudflare says makes the setup less exposed to hidden instructions in incoming text.

Cloudflare's own numbers

In Cloudflare's own test, this chain took 19.3 seconds end to end. The same task took 61.7 seconds when the language model GLM-5.3 handled all three steps. Cloudflare puts the cost at 6,032 dollars per million cases for the chain, against 11,740 dollars when GLM-5.3 did everything. That is 49 percent lower. The figures are the company's own, and the test has not been independently verified.

Cloudflare also tested forecasting on public figures for the number of travellers passing through security checkpoints at US airports. Over 21 days the model was off by 3.3 percent on average, and 17 of the 21 actual values fell inside its interval.

What Cloudflare itself warns about

The announcement also lists the limitations. Cloudflare writes that reading works well, but that arithmetic over many dependent steps is not yet reliable. The same applies to aggregating across many rows of a table. The company recommends letting the model extract the values and letting ordinary code do the maths.

Estimates require the document to contain comparable figures or history. Calibration of the intervals is measured, not guaranteed, and can run narrow on unfamiliar data. The forecasting part is described as basic and not suited to detailed long-horizon forecasts. Cloudflare writes that every answer is always well formed, but not necessarily correct.

The announcement says nothing about where data is processed, how long it is stored, or which contract terms apply to the preview.

What leaders should note

• Finance, procurement and customer service teams that use language models to read invoices, receipts and complaints can test Clairic as a cheaper first step. The gain has to be measured on their own documents, not only on Cloudflare's examples.
• Traceability through character positions and confidence scores makes it easier to document why a figure was accepted. That can simplify audit and internal control, but only if thresholds for manual review are set and followed up.
• This is a preview on Cloudflare's platform. Before personal or accounting data is sent in, the data processing agreement, processing location and retention period should be clarified, since the announcement does not state them.
• Open weights could later make it possible to run the model in-house and reduce vendor dependency. Until the weights are actually released, that remains an intention on Cloudflare's part.

Sources and media

• Cloudflare: Introducing Clairic, a new System One class of extraction models, blog post published 19.10.2026 at 19:48 Oslo time
• Sources were checked on 20.10.2026.
• Illustration: AI-generated with Grok Bot / hogby.ai