Cloudflare releases open weights for its Clairic document extraction models
Cloudflare on November 18 released open weights for a redesigned Clairic, a 27-billion-parameter model that pulls typed fields out of text, images and PDFs, and a new 9-billion-parameter Clairic-flash. Both run on Workers AI, and Clairic's price rises to $0.24 per million input tokens after December 1.
Cortex, covering AI models, labs and agents
, 3 min read
$0.24Clairic price per million input tokens after Dec 1
What to know
- Cloudflare published open weights for a redesigned 27B Clairic and a new 9B Clairic-flash, document extraction models that now read images and PDFs as well as text.
- On Workers AI both cost $0.10 per million input tokens until December 1; Clairic then moves to $0.24 while Clairic-flash stays at $0.10. Output tokens are free.
- The models come from a new Workers AI Superintelligence team, which Cloudflare says also built its open-weight Clef decision models released October 1.
A content delivery network now has a model research team with "superintelligence" in its name. That's the part of Cloudflare's November 18 post that will get the eye-rolls. The part worth reading is narrower and more useful: two small models that take a receipt, an email or a scanned PDF and hand back typed fields, with a confidence score on each one, and whose weights anyone can now download.
Clairic, first shown as a preview on October 19, has been rebuilt. The name now belongs to a 27-billion-parameter model, and a new 9-billion-parameter Clairic-flash takes the lighter slot. Both now read images and PDF pages, up to four per request, which the preview could not do with PDFs. Line items come back as rows. Cloudflare says the output schemas are unchanged, so preview users were moved to the new model automatically. Weights are on Hugging Face under Cloudflare's account, and the hosted versions sit on Workers AI. The post calls them open-weight but, in what we could read, does not name a license; Cloudflare's earlier Clef models shipped under Apache 2.0, according to The Register.
Cheap, then less cheap
Through December 1 both models cost $0.10 per million input tokens, and output is free. After that Clairic goes to $0.24 and Clairic-flash stays at $0.10, prices RuntimeWire also reports. That is a 2.4x increase for the bigger model, but the base is tiny. A 1,000-token invoice costs $0.00024 on the new Clairic rate, or $240 for a million of them. For a back office running document intake, the flash model at less than half the price is the obvious first test.
Charging only for input is a sensible fit for the job. An extraction model's output is a handful of fields, so input is nearly the whole bill anyway. See our explainer on per-million-token pricing for why that framing matters.
Whose benchmark
Cloudflare's numbers are its own, and to its credit it says so, calling them directional because they are self-reported. On a synthetic set of 576 receipt fields it reports the new Clairic and Clairic-flash both got all 576 right, against 369 for the preview, or 64%. Its comparison models were Liquid AI's D1.2 paired with Apple Vision for text recognition, and ScaleDown's Extract, both run by Cloudflare. A perfect score on a set the vendor built is a claim about fit to that set. Nobody outside Cloudflare has published a test yet.
The design choice is the more interesting technical detail. The preview filled every field in one parallel pass. The new version generates fields in parallel slots, each stopping at its own end marker, a halfway house between one-shot extraction and ordinary token-by-token generation. The October preview notes, as Unite.AI reported them, flagged multi-step arithmetic and sums across many table rows as unreliable. The new post does not say whether that limit is gone.
Why a CDN builds models
Clairic is half of a pairing. Clef, the open-weight decision models released October 1, picks an action; Clairic pulls the facts from the document first. That is the same split Google drew with Gemini Decision 1, and it is the plumbing that agents like Cognition's Devin Bots need underneath them. Cloudflare's angle is that small, specialized, open-weight models are cheap to run at the network edge it already owns. A superintelligence team is a big name for two models under 30 billion parameters. Whether it fits depends on what that team ships next, and whether outsiders' tests match Cloudflare's.
Sources
- An updated Clairic and Clairic-flash are now open-weight, from the new Workers AI Superintelligence team, Cloudflare, November 18, 2026
- Cloudflare releases open weights for two document extraction models, RuntimeWire, November 18, 2026
- Cloudflare releases Clairic, its first System One extraction model, Unite.AI, October 19, 2026
- Cloudflare tries to outplay Jev with open-weight Clef models, The Register, October 1, 2026
- Clairic, OpenRouter
AI modelsAI agentsAI chips and data centers
Written by an AI correspondent under standards set by editor Jacob Knight. Spotted an error? Tell us and we will fix it.