Cloudflare previews Clairic, a System One extraction model that returns values instead of choices
Cloudflare announced Clairic on 19 October 2026 (blog, Alex Reneau, Michelle Chen and Kevin Flansburg; X, 15:06 BRT). It calls it the first of a new class, System One extraction models: where Clef and Jev pick among options you list, Clairic returns values pulled from the input, computed from it or forecast beyond it.
What it does. You send a state and typed fields (text, date, number with decimals and unit, list, and estimate with a min, max and step). Text answers come back as exact spans with character offsets; a missing value comes back as null; estimates come back as a median with an 80% interval (p10 to p90). It is on Workers AI as @cf/cloudflare/clairic, in preview, with vision, a 65,536-token context and $0.10 per million input tokens (docs). Cloudflare says it intends to release open weights “soon”; there is no repository yet.
Vendor numbers (UnverifiedClaim). In a support-email cascade (Clairic extracts, Clef decides, GLM-5.3 writes the reply), Cloudflare reports 19.3 s end to end against 61.7 s for GLM-5.3 doing every step, and 49% lower cost. On 21 days of TSA checkpoint counts, it reports forecasts off by 3.3% on average, with 17 of 21 actual values inside the 80% interval.
Why it matters, and why there is no catalog page. It extends the typed, probability-returning API style from choices to values. Our catalog covers decision models (choice, score, boolean), and Clairic answers a different kind of question, so whether it gets a page is an open scope decision; for now it is news only.
How it was framed on X (20 Oct)
Kevin Flansburg of Cloudflare described Clairic (19 Oct, 15:24 BRT) as “a novel system one model … based on regression instead of classification (i.e. decision models like Clef)”, for continuous values without listing choices upfront, and repeated “Open weights coming soon!”. Morgan Linton quoted it (19 Oct, 17:12 BRT), arguing that “most decision models are still LLMs at their core, not System One models like Jev”. That is his opinion, not a measured claim; Cloudflare has published no Clairic results on a public board.
AIE CODE, 11 Nov 2026
Michelle Chen walked through Clairic on the AIE CODE Day 1 mainstage (“cloudflare is a neolab”, 9:50–10:10 PT). Livestream, about 0:48–0:59:52. No official captions; this note follows an automatic transcript, which wrote “Cleric”. Full recap: AIE CODE Day 1.
She described the same shape as the 19 Oct post: one forward pass, unstructured input (email, tweet, Slack, or an image), typed JSON out (extract, compute, estimate, forecast). She said Clairic was an early preview and that a new launch is coming out. That is not GA, and it is not open weights shipping on 11 Nov.
Demo figures she stated (UnverifiedClaim). In a CI-triage sketch, Clef Flash prunes an agent’s tools (10 of 40) and Clairic pulls a PR number and an error from 300–600 log lines, then one LLM acts. Live: 6 LLM calls vs 1, and 40,000 LLM tokens vs 1,800. Prepared offline run, not conference Wi-Fi: 9.2 s vs 17.2 s, 72% less expensive, calls from five down to one, about 94% fewer tokens. The live line said six; the offline recap said five. Slides also showed Clairic on web pages and support tickets. None of this is an independent benchmark.
Still no catalog page. The catalog covers decision models (choice, score, boolean). Clairic returns values. Whether it gets a page stays an open scope decision. The 11 Nov talk does not put Clairic on the quadrant.
