On Wednesday Cloudflare released open weights(opens in a new tab) for Clairic and a smaller Clairic-flash, two models built for one job: pulling named fields out of text, images and PDFs. You ask for an order number, a date, an amount paid or a tracking number, and you get back typed values, each with a confidence score, or a null when the value is not in the document. It does not write prose. Clairic-flash has 9 billion parameters; Clairic is the 27-billion-parameter version, Unite.AI reports(opens in a new tab). Open weights means anyone can download and run the model on their own hardware.
The price is the news for anyone who copies data out of documents for a living or pays someone to. On Cloudflare's hosted service, Clairic costs $0.24 per million input tokens and Clairic-flash $0.10, with no charge for output, and both cost $0.10 until December 1. A token is roughly three quarters of a word. If a one-page receipt runs about 1,000 tokens, an assumption for illustration, a million receipts would cost about $240 at the standard rate. In its October preview post(opens in a new tab), Cloudflare worked through a damaged-product claim and put the cost at $6,032 per million claims when Clairic did the extraction, against $11,740 when a general model did every step.
Treat the accuracy claims with care. Cloudflare says plainly that no public extraction benchmark exists yet and that its tables are self-reported. Its tests show perfect scores on synthetic receipts and on its customers' own checks, and a few misses on fresh text. That is a reason to test, not a reason to switch.
If you or your team pull fields out of invoices, receipts, claim emails or order forms, this week take 50 documents you have already checked by hand and run them through Clairic-flash while the trial price lasts. Compare every field, count the wrong values and the missed nulls, and see whether the low-confidence scores land on the fields that were wrong. If they do, send only those fields to a person.
Action items
On Wednesday Cloudflare released open weights for Clairic and Clairic-flash, two models that do one job: pull named fields out of text, images and PDFs, each with a confidence score, or a null when the value is missing. Hosted, they cost $0.24 and $0.10 per million input tokens, with no output charge, and both cost $0.10 until December 1.
Cloudflare's accuracy tables are self-reported, and it says no public extraction benchmark exists yet. Test before you switch.
If you pull fields out of invoices, receipts, claim emails or order forms, run 50 documents you have already checked by hand through Clairic-flash this week. Count the wrong values and missed nulls, check whether low confidence lands on the wrong fields, and if it does, send only those fields to a person.
Researched and drafted by an automated workflow, then reviewed and edited by a human editor before publication. Every source is linked. See how we use AI here.
Also worth knowing
- The Cloudflare BlogAn updated Clairic and Clairic-flash are now open-weight(opens in a new tab)
Two extraction models with open weights, billed on input only: $0.24 and $0.10 per million tokens, both $0.10 until December 1. Each accepts up to four images or PDF pages per request. The accuracy tables are Cloudflare's own.
- TechCrunchMeta's AI agent Muse is mixing up conversations and could be exposing private data, report says(opens in a new tab)
Meta employees testing Muse saw it volunteer bank balances they never shared and remind them about Signal messages. Meta told Reuters many were hallucinations and called the incidents isolated. Before granting an agent notification or screen access, decide what it should never see.
- New York State Attorney GeneralAttorney General James announces creation of AI Safety, Standards, Enforcement, and Transparency Unit(opens in a new tab)
New York now has a dedicated unit to enforce the RAISE Act, which takes effect January 1 and makes large frontier model developers disclose safety measures and report incidents. If you buy from those developers, ask what they will report and when.
- Office of the Governor of CaliforniaCalifornia releases next steps on mandatory AI safety standards and oversight(opens in a new tab)
Newsom's report proposes baseline standards for all AI companies, independent testing, and state power to order an emergency model shutdown. These are recommendations with no dates or penalties yet, so track them; there is nothing to comply with today.