Introducing Clairic, a new System One class of extraction models

Open original

Cloudflare is previewing Clairic, a model that pulls values out of unstructured text, computes them, or forecasts them in a single pass with no generated tokens. It's pitched as faster, cheaper, and less error-prone than using an LLM for the same structured-output jobs.

  • Clairic returns typed values (spans, numbers, dates, durations, lists, distributions) rather than free text, so it can't produce a malformed or hallucinated answer; missing values come back as null.
  • Every field is answered in one parallel pass with no output tokens billed, so asking for thirty fields costs about the same as asking for three, and responses usually land in under a second.
  • Each answer carries confidence scores, character-offset spans back to the source, and 80% intervals for estimates, so you can see which outputs to double-check.

Cloudflare is previewing Clairic, the first model in a new category it calls "System One extraction models." The idea builds on TypeSafe's System One framing, where a model is fast, narrow, and gives calibrated answers in a single pass without generating reasoning tokens. TypeSafe's Jev and Clef are decision models: you give them a list of options and they return probabilities over those options. Clairic instead pulls out values. It reads unstructured input and returns extracted text spans with character offsets, typed values like numbers, dates, durations and lists, estimates bounded by ranges you declare, and forecasts for future steps in a series. Every scalar answer comes with a null probability and a confidence score, and estimates come back as a median plus an 80% interval. Cloudflare says it is faster and cheaper than frontier LLMs for these jobs, charges only for input tokens, and typically responds in under a second. The practical pitch is aimed at the large volume of structured-output calls that teams currently send to general LLMs: invoice totals, delivery dates, how many days late a shipment is, expected repair costs, next-day traffic. These questions have no answer list you can enumerate in advance, so decision models don't fit. Clairic's argument is that LLMs are overkill and error-prone here. Generated output can be malformed JSON, a digit can be misread mid-generation, and a pulled-out excerpt can't be traced back to its source without extra work. Clairic avoids those failure modes by construction, since it outputs values of the declared type (or null) and grounds extracted text to specific positions in the input. Cloudflare frames the three model types as a pipeline: extraction turns raw input into typed values, decision models turn values into choices, and LLMs turn choices into prose. The claims are mostly architectural at this stage, and the post is light on benchmarks in the excerpt provided, so the accuracy and cost comparisons against LLMs are worth verifying before anyone rebuilds a pipeline around them. The model is available now as a hosted preview on Workers AI under `@cf/cloudflare/clairic`, with open weights promised but not yet released. For anyone running high-volume extraction through a frontier model today, the interesting question is whether Clairic's error rates and per-call cost hold up on their own documents, not on Cloudflare's examples.

Ask about this article