Cloudflare has released a preview of Clairic, a new System One extraction model on Workers AI. The Cloudflare Clairic extraction model reads text or images and returns typed values, not generated text. It targets a job developers usually give to LLMs: pulling totals, dates and IDs into JSON. Clairic answers every requested field in 1 forward pass, with no reasoning and no output tokens.
TL;DR
- Size: Parameter count not disclosed. 65,536-token context window, up to 4 images per request.
- Runs on: Cloudflare-hosted GPUs via Workers AI (
@cf/cloudflare/clairic). No local option until weights ship. - Performance: Sub-second extraction in Cloudflare’s tests. No standard accuracy benchmark is published yet.
- Best: Extraction step took 0.46 s vs 16.1 s for GLM-5.3 in Cloudflare’s claims pipeline.
- Worst: 4 of 21 TSA next-day forecasts fell outside its 80% interval. Multi-step arithmetic is unreliable.
What is Clairic?
Clairic is a Cloudflare model that extracts, computes, estimates and forecasts typed values from unstructured inputs. It builds on the System One idea from TypeSafe’s Jev, named for Kahneman’s fast, intuitive thinking.
Jev and Cloudflare’s own Clef are decision models. They score options you list in advance. Clairic covers answers you cannot list, such as an invoice total, a date or a repair cost.
Cloudflare frames a 3-part stack. Extraction models turn raw input into typed values. Decision models turn values into decisions. LLMs turn decisions into words.
How does a Clairic request work?
A request has a state (text plus up to 4 images) and a map of named, typed fields. Each field can carry plain-language instructions. Supported types are text, number, date, time, datetime, duration, list, estimate and forecast, per the Workers AI changelog.
Every scalar answer includes p_null, the probability that the value is absent, plus a confidence score. Text answers return exact spans with character offsets. Estimates and forecasts return a median and an 80% interval, from p10 to p90. Forecasts cover up to 64 steps.
There is no field limit. Fields that do not fit beside the input in one 8,192-token pass split into parallel passes. Each pass sees the whole input. In the docs example, a 61-token invoice ran in 3 passes. It billed 655 input tokens, including 472 schema tokens. Wide schemas can cost more than the document itself.
What did Cloudflare’s tests show?
In a support-email example, Clairic added price and tax into $701.10. It also computed 11 days since delivery. Neither number appears in the email. A missing tracking number returned null instead of a plausible guess.
On a skewed receipt photo, it read all 7 present values and returned null for the absent discount. Its lowest confidence was 0.776, on the order number.
For forecasting, Cloudflare predicted daily TSA checkpoint counts for 21 days. Average error was 3.3%. 17 of 21 actual counts fell inside the 80% interval. 3 of the 4 misses fell on Columbus Day weekend.
The main test is a damaged-product claims pipeline. Clairic extracted facts, Clef chose refund or repair, and GLM-5.3 wrote the reply. Cloudflare compared this against GLM-5.3 handling all 3 steps.
