Cloudflare's Clairic pulls typed values out of messy text, and charges only for input

Clairic is a new model from Cloudflare that reads unstructured input, such as an email, an invoice or a photo of a receipt, and returns the specific values you ask for. Cloudflare calls it a "System One" extraction model: it answers in a single pass without reasoning or writing text. It is out today as a preview on Workers AI under the id @cf/cloudflare/clairic, and Cloudflare says it intends to release the weights openly "soon."

What it does. Cloudflare lists four jobs. Extract a value that is in the input, like an order ID. Compute one from facts in the input, like a total including tax or the days since delivery. Estimate a value the input only gives evidence for, like a sale price from comparable sales. Forecast the next steps of a series. Text answers come back as spans with character offsets, so you can point to where they appeared in the source. Numbers and dates follow their declared type, and every single-value answer carries a confidence score and a p_null, the model's probability that the value isn't in the input. If it can't find a value, it returns null rather than a plausible-looking guess. Estimates and forecasts return a median and an 80% interval.

How you use it. A request is a state (the text, plus up to four PNG, JPEG or WebP images, which must be embedded because remote URLs aren't accepted) and a map of named fields, each with a type (text, number, date, time, datetime, duration, list, estimate, forecast) and optional plain-language instructions. You call it with env.AI.run() from a Worker or through the Workers AI REST API. The docs list a 65,536-token context window and a price of $0.10 per million input tokens. Because it generates no tokens, there is no output charge. There is no cap on the number of fields; fields that don't fit next to the input in one pass are split across parallel passes.

Cloudflare's own numbers. In one test, Cloudflare asked Clairic to forecast the next day's TSA airport checkpoint count for 21 days in a row. It says the forecasts were off by 3.3% on average and 17 of the 21 actual counts landed inside the 80% interval, with most misses on the Columbus Day weekend. In a damaged-product claim demo, a pipeline of Clairic for extraction, Cloudflare's Clef decision model for the decision and GLM-5.3 for the reply took 19.3 seconds end to end and cost $6,032 per million claims. Using GLM-5.3 for every step took 61.7 seconds and cost $11,740. These are Cloudflare's tests, not ours.

Who it's for: developers on Cloudflare who send the same structured-output calls to a large language model over and over, such as "what's the invoice total, the due date, the order number," and want typed answers with confidence scores they can check before automating anything.

The caveat, from Cloudflare itself. The hosted endpoint is "an early checkpoint." Cloudflare says one- and two-step calculations seem reliable, but longer chains of dependent arithmetic and adding up many rows of a table are not yet, and its advice is to "have Clairic predict the value and let code do the math." It also says the intervals can run narrow on unfamiliar data and that forecasting is "basic," not a replacement for a dedicated time-series model. The docs add that typed output does not guarantee correctness: an answer can have the right shape and still be wrong. The model is offered "as is" in preview, and Cloudflare says it may stop providing it.

We haven't tested Clairic ourselves; this is based on Cloudflare's announcement and docs, read on October 19. Not sponsored, no affiliate links.