Cloudflare releases Clairic to extract data without generating text

Cloudflare introduced Clairic on October 19, 2026, to pull usable values from text and images without generating a written response. The preview model is available on Workers AI. Cloudflare says it typically answers in under a second, but warns that well-formed results are not necessarily correct—and longer calculations remain difficult.

Developers can call the preview as @cf/cloudflare/clairic from a Cloudflare Worker using env.AI.run(), or through the REST API. Its listed context window is 65,536 tokens, with pricing of $0.10 per million input tokens. There is no output-token bill because the model does not generate output tokens.

Values instead of a written answer

A request supplies text, up to four optional images, and a set of fields defining the desired answers. Those fields can ask for text, numbers, dates, times, durations, lists, estimates or forecasts, with plain-language instructions attached. Cloudflare calls this a System One extraction model.

Cloudflare says the model answers fields in parallel, without a separate reasoning step. Field sets too large for one pass are divided across parallel passes, each seeing the full input, then merged. Individual values carry confidence scores and p_null, a probability that the requested value is absent. An undetermined field returns null, alongside a separate guess.

Cloudflare’s examples

• Support email: Clairic added a $649 price and $52.10 tax to return $701.10, calculated an 11-day interval between delivery and the message, and returned null for an unstated tracking number.
• Receipt photo: It extracted the order number, purchase date and time, quantity, unit price, subtotal, sales tax and tip. A discount absent from the receipt came back as null.

A claims workflow with fewer language-model steps

The company’s damaged-product claims test used Clairic for extraction, Clef for the decision, and GLM-5.3 for the generated response. It compared that chain with a version using GLM-5.3 throughout. Clairic and Clef handled the first two stages in under a second combined, versus 38 seconds for the language model, according to Cloudflare.

Forecasting received a separate, narrower test. Over 21 days, Cloudflare used Clairic to predict next-day TSA airport checkpoint counts. It reported an average error of 3.3%, with 17 actual counts falling inside the model’s 80% prediction interval. Cloudflare calls forecasting basic, not a replacement for dedicated time-series models.

Where the shortcut runs out

Cloudflare says one- and two-step arithmetic seems reliable, but longer chains of dependent calculations and sums across many table rows remain difficult. The one-pass model lacks a scratchpad for working through those steps. Its advice is to use Clairic for prediction and handle arithmetic in code.

Estimates return a distribution within caller-defined bounds, including a median and an 80% interval. They need supporting evidence in the input, and intervals can be too narrow on unfamiliar data. Cloudflare advises using confidence scores, missing-value probabilities and intervals to decide which results to automate or review. It plans further checkpoint work before releasing Clairic’s weights, with no date set.