Intel / blog

Introducing Clairic, a new System One class of extraction models

Source: blog.cloudflare.com
Date: Published Oct 19, 2026

Context

Cloudflare's Clairic is a model for pulling values out of messy input, such as an invoice total, a delivery date or a likely repair cost, that answers in one fast pass instead of generating text the way a does (blog.cloudflare.com). It is available as a preview on Workers AI, and Cloudflare intends to release soon. Cloudflare places it in the System One category that TypeSafe introduced with Jev: models meant to be fast and precise, as opposed to System Two models such as LLMs that reason deliberately. Until now those models, Jev and Cloudflare's own Clef, have been decision models: they take an input and a list of options and return a probability for each. Clairic covers questions whose answers cannot be listed in advance.

A request is the input text, plus up to four images, and a set of named fields, each with a declared type such as text, number, date, duration, list, estimate or forecast. The model answers every field in a single forward pass, generates no , and charges only for input. Text answers come back as spans with their character positions in the source, numbers and dates follow the rules of their type, and estimates and forecasts come back as a median with an 80% interval inside bounds the caller sets. Each answer is a value of the requested type or null, with a confidence score and a probability that the value is absent. In Cloudflare's support-email example, Clairic added the price and tax to report the amount paid, worked out the days since delivery from two dates, and returned null for a tracking number the email did not contain. Cloudflare notes the model cannot produce a mistyped answer but can still be wrong.

Cloudflare's main evidence is a damage-claim pipeline in which Clairic extracts the facts, Clef chooses refund or repair, and GLM-5.3 writes the reply. Against GLM-5.3 doing all three steps, it finished in 19.3 seconds instead of 61.7 and cost $6,032 per million claims instead of $11,740, and Cloudflare says both versions got the facts and the decision right in every run. In a 21-day test forecasting daily TSA checkpoint counts, forecasts were off by 3.3% on average and 17 of 21 actual counts fell inside the 80% interval, with three of the misses on the Columbus Day weekend. Cloudflare calls the endpoint an early checkpoint: one- and two-step arithmetic seems reliable but longer chains do not, its intervals can run narrow on unfamiliar data, and its forecasting is basic and not a replacement for a dedicated time-series model.

Key takeaways · AI-distilled

Cloudflare's Clairic, in preview on Workers AI, answers typed fields (spans with character offsets, numbers, dates, durations, lists, estimates and forecasts) in one forward pass with no output tokens, so it bills only input and typically responds in under a second.
Each Clairic answer is a value of its declared type or null, with p_null and a confidence score, and estimates return a median and 80% interval inside declared bounds. In Cloudflare's example, a missing tracking number came back null instead of a guess.
In Cloudflare's damage-claim test, a Clairic, Clef and GLM-5.3 cascade took 19.3s versus 61.7s with GLM-5.3 doing every step, and cost $6,032 versus $11,740 per million claims. Clairic's repair estimate stayed identical across runs; GLM-5.3's changed.
On 21 days of next-day TSA screening forecasts, Clairic averaged 3.3% error with 17 of 21 actual counts inside its 80% interval. Three of the four misses fell on Columbus Day weekend, an event the series itself does not show.
Cloudflare says this early checkpoint reads well but is not reliable on long chains of arithmetic or aggregation across many table rows, so code should do the math, and its intervals can run narrow on unfamiliar data. Open weights are planned.

Terms in this piece · Glossary

LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.

Read the source blog.cloudflare.com