An updated Clairic and Clairic-flash are now open-weight, from the new Workers AI Superintelligence team
Cloudflare's Workers AI team has open-sourced an updated Clairic model and a smaller 9B version called Clairic-flash, both built for pulling structured data out of text, images, and PDFs. They claim strong accuracy and sub-second response times at a low per-token price.
- The redesigned architecture fills fields in parallel with autoregressive generation, lifting accuracy on text and images; image-reading errors fell from 207 to 0 on their internal benchmark.
- Clairic now reads PDFs and images (up to four pages or images per request), so supplier, line items, and totals can be pulled in one call even across pages.
- Pricing is $0.24 per million input tokens for the 27B Clairic and $0.10 for the 9B Clairic-flash, with a promotional $0.10 rate on Clairic until December 1, 2026. Most evals are self-reported, which the authors acknowledge.
Cloudflare has released an updated version of Clairic, its extraction model, along with a smaller sibling called Clairic-flash, and both are now open-weight on Hugging Face and hosted on Workers AI. The 27B Clairic and 9B Clairic-flash are the two sizes. The main change from the preview is architectural: ordinary text and number fields now use parallel slot autoregressive generation rather than a single forward pass, with each field generating its own tokens until its own end marker, all advancing together in the serving batch. Cloudflare also trained special heads for estimates, forecasts, durations, calculated dates, and arithmetic. The model now reads PDFs and images, not just photos of receipts, and can pull supplier, line items, and totals from up to four images or PDF pages in one request. Confidence scores, character spans, and probabilities are still returned with each answer.
The numbers Cloudflare reports are self-described and come from internal evals, so they deserve some skepticism. On a set of 336 fresh text fields, the updated Clairic got 335 right, Clairic-flash got 328, and the preview got 295. On receipt-style fields in text and images, the new model hit 576 of 576 where the preview managed 369. Against Liquid AI's D1.2 pipelined with Apple Vision OCR and ScaleDown's Extract, Clairic comes out well ahead on Cloudflare's own tests. Latency numbers are more concrete: p50 times run from about 155 ms for Clairic-flash on fresh text fields up to roughly 630 ms for Clairic on expense audits, with p95 values mostly under a second. Cloudflare admits there are no public extraction benchmarks yet, so independent verification is the thing to watch for.
On price, the 27B Clairic lists at $0.24 per million input tokens and Clairic-flash at $0.10, with output tokens not billed. Until December 1, 2026, both cost $0.10 per million input tokens, so anyone on the preview endpoint gets the better model automatically at the old price. After that date the discount ends and Clairic goes back to $0.24, which means cost-sensitive workloads should probably be moved to Clairic-flash before then. The pitch overall is that these are narrow, cheap models meant to sit alongside expensive general LLM calls for extraction work, and the open weights are an invitation to people to fine-tune them for their own tasks.
Ask about this article
is this offered with a BAA for hipaa?
Short answer: Cloudflare signs BAAs, but only with Enterprise customers. Whether Clairic specifically is covered is less clear.
The article itself says nothing about HIPAA. It does say both models are [hosted on Workers AI](https://blog.cloudflare.com/open-clairic/?utm_source=tldrai), and Workers AI is listed in Cloudflare's HIPAA scope. Cloudflare's [HIPAA FAQ](https://www.cloudflare.com/trust-hub/compliance-resources/hipaa/) names Workers AI among the in-scope Developer Platform services. So the platform is covered, but that doesn't automatically extend to a model that launched in this release.
On the BAA itself, Cloudflare's [Trust Hub page](https://www.cloudflare.com/trust-hub/compliance-resources/hipaa/) says it enters into a BAA with Enterprise-level customers whose use of our services involves PHI. The [US privacy compliance page](https://www.cloudflare.com/trust-hub/us-privacy-compliance/) puts it more bluntly: Cloudflare will only enter into business associate agreements (BAAs) with its enterprise customers. The BAA also must be formally executed by both parties before any PHI is processed through the Cloudflare network.
Some third-party summaries claim self-serve or lower-tier customers can get one, but those sites contradict Cloudflare's own wording, so I'd trust the Trust Hub pages. Also note that Cloudflare's AoC covers in-scope services, and a BAA only applies to the services it names, so confirm with sales that Workers AI and these specific models are included.
Practical next steps:
- Contact Cloudflare sales through their [enterprise contact page](https://www.cloudflare.com/plans/enterprise/contact/) and ask whether the BAA covers Workers AI and the Clairic endpoints.
- Existing customers can download the HIPAA Attestation of Compliance from the dashboard's Resources section; prospective customers request it under an NDA from their account executive.
- Don't send PHI to Clairic until the BAA is signed. Cloudflare's own guidance is that use of Cloudflare services for PHI without a signed BAA may result in non-compliance for your organization.