Cloudflare has opened a preview of Clairic, a model that reads an email, receipt or spreadsheet and hands back the figures inside it instead of writing a sentence about them. It runs only on Cloudflare’s Workers AI platform, as @cf/cloudflare/clairic. Open weights are promised “soon,” according to the Cloudflare blog post, and the company gave no date.

A developer sends the text, plus up to four images, along with a list of named fields. Each field gets a type: a name, an amount, a date, a duration, a list, a range estimate or a forecast. Clairic answers every field in one pass, generates no output text, and bills only for input. Cloudflare says responses typically arrive in under a second.

The company’s claim about errors is narrow, and the wording matters. Because the model emits a value rather than composing text, Cloudflare says it cannot return a mistyped answer or hallucinate in the way an LLM does. The same sentence concedes it can still make mistakes. Later the post goes further: the design guarantees a well-formed answer, and a well-formed answer can still be wrong.

The more useful product decision is the scorecard attached to each answer. Every scalar value arrives with a confidence score and a probability that the value is absent from the input. Estimates arrive with a median and an 80 percent interval. In Cloudflare’s own support-email demo, the model was asked for a tracking number the email never contained. It returned null, and its closest guess, a tax amount, carried a confidence of about 0.42. A pipeline can route anything under a chosen threshold to a person.

The category is not Cloudflare’s invention. Cloudflare credits TypeSafe, which announced “System One” models about a month earlier with a model called Jev, and credits Daniel Kahneman’s book Thinking, Fast and Slow for the fast-versus-deliberate framing. Jev and Cloudflare’s own Clef pick among options a developer lists. Cloudflare’s addition is a second branch, which it calls extraction, where the answer is a value read from the input, worked out from it, or predicted beyond it.

The post’s headline claim of higher accuracy, lower latency and lower cost than LLMs rests on one test. Cloudflare ran a damaged-product claim through Clairic, then Clef, then GLM-5.3 to write the reply, and compared that with GLM-5.3 doing all three steps. The mixed pipeline took 19.3 seconds against 61.7, and cost $6,032 per million claims against $11,740. Accuracy did not separate the two: every run of both versions reached the correct facts and the correct decision, so what the test shows is a gap in speed, cost and consistency. Across repeated runs, Clairic produced the same repair figure every time, whereas GLM-5.3 drifted: $250, $250, then $220. Cloudflare also says Gemma 4 and Qwen could not deduce the repair cost at all. These are the company’s own measurements on its own task, with no independent evaluation.

Forecasting gets a smaller claim. Asked to predict the next day’s U.S. airport screening count from public TSA data, Clairic was off by 3.3 percent on average over 21 days, and 17 of the 21 actual counts fell inside its 80 percent interval. Three of the four misses landed on the Columbus Day weekend, which the data alone could not signal. Cloudflare says plainly that this is no replacement for a dedicated time-series model.

The company lists other weak spots in this early checkpoint. Chains of dependent arithmetic steps are unreliable, so its advice is to have the model read the values and let ordinary code do the math. Intervals can run too narrow on unfamiliar data, and estimates need comparables or history in the input to work from.

For teams paying frontier-LLM rates to pull fields out of invoices and receipts, the cheap test is to run a month of real documents through both and set the confidence cutoff from the misses. Cloudflare has not said when the weights will ship, and until it does, any savings come with a Workers AI dependency.

Cloudflare, “Clairic” announcement on the Cloudflare blog, published 19 October 2026.