Introducing Clairic, a New System One Class of Extraction Models

Cloudflare has introduced Clairic, a preview of what it calls a new class of System One extraction models. The first version is already hosted on Workers AI as @cf/cloudflare/clairic, and the company says open-weights are coming soon.

The pitch is simple: stop asking a general-purpose model to do a very specific job badly. Clairic is built to extract, compute, estimate, and forecast values from unstructured input, then return typed answers instead of free-form text. Cloudflare says that means better accuracy, lower latency, and lower cost than traditional LLMs for the same kind of work.

This comes after TypeSafe’s System One models, which Cloudflare says helped kick off the idea of a separate model class for fast, precise tasks. TypeSafe’s Jev and Clef are decision models: they take input and pick among options. Clairic is different. It is for extraction, not decision-making, and it answers in one pass without reasoning or token-by-token generation.

That matters because so much of today’s model usage is still boring data plumbing. Invoice totals, delivery dates, late days, repair estimates, traffic forecasts. Cloudflare argues those jobs do not fit the usual LLM pattern, especially when the output needs to be a number, a date, a list, or a distribution rather than an essay. Clairic is aimed squarely at that gap.

The model returns spans, numbers, dates, durations, lists, estimates, and forecasts, along with confidence scores and a probability that the answer is missing. Cloudflare says it can also work from images, not just text, and that it can split large field sets into parallel passes when needed. The bigger point is that the output is structured by design, so the model is supposed to give you something a workflow can actually use instead of a plausible-sounding paragraph.