Cloudflare has put a preview of Clairic on Workers AI, a model built to pull totals, dates and identifiers out of unstructured input and return them as typed values.
It answers every requested field in one forward pass. There is no reasoning step and no output tokens, which is the design choice that separates it from the general-purpose assistants developers usually reach for.
Cloudflare says the context window reaches 65,536 tokens and requests can carry up to four images. Supported types include text, number, date, time, datetime, duration, list, estimate and forecast, and each field accepts plain-language instructions.
The stack splits into three jobs
Clairic rounds out a division of labour the company has been building. Extraction models turn raw input into values. Decision models turn values into decisions, which is the terrain of Cloudflare’s earlier Clef pair and TypeSafe’s Jev. Language models turn decisions into sentences.
Speed is the headline claim. Cloudflare reports an extraction step finishing in 0.46 seconds against 16.1 seconds for GLM-5.3 in its own pipeline. No standard accuracy benchmark has been published.
The company’s own examples show the limit. Four of 21 next-day TSA forecasts landed outside an 80% interval, and multi-step arithmetic stays unreliable.
For teams paying per token to parse invoices, the appeal is obvious. Most of that spend buys fluent prose nobody reads, and a model that returns a filled form skips the whole detour.
