Cloudflare open-weights Clairic, a 27-billion-parameter extractor, and prices the 9-billion flash model at ten cents
Alex Reneau, Michelle Chen and Kevin Flansburg’s post is dated 18 November. Clairic, 27 billion parameters, moves to the old preview endpoint. Clairic-flash, 9 billion, is the new lighter model. Both are on Hugging Face and on Workers AI. Until 1 December both cost $0.10 per million input tokens, and Cloudflare says it does not charge for output tokens. The scores are the company’s. On one receipt set the preview missed 207 of 576 fields and the new models missed none.
Feature3 minIn-depth
James Okonkwo3 min read
- Cloudflare
- Clairic
- Open weights
- Extraction

Photo: CloudflareSource page
Cloudflare on Wednesday released the weights for two extraction models and moved the larger one onto the endpoint that had been serving the preview. Alex Reneau, Michelle Chen and Kevin Flansburg’s post is dated 18 November. Clairic is the 27-billion-parameter model. Clairic-flash is the 9-billion-parameter model. Both are hosted on Workers AI, Clairic at `@cf/cloudflare/clairic` and Clairic-flash at `@cf/cloudflare/clairic-flash`, and the weights are on Hugging Face. The preview endpoint of that first name had been the 9-billion-parameter checkpoint, priced at $0.10 per million input tokens. The post says the 9-billion model now lives at the flash name, and the 27-billion model takes the old name. An application still pointed at the preview endpoint gets the larger model as the change rolls out.
The list price after the promotion is $0.24 per million input tokens for the 27-billion model and $0.10 for the 9-billion model. Until 1 December, the post says, both cost $0.10. Cloudflare says the models do generate output tokens and that it does not charge for them: “we eat the cost of the output token generation.” The architectural change described for ordinary text and number fields is parallel slot autoregressive generation, instead of a single forward pass. Each field writes its own tokens and stops at its own end marker, and the fields advance together in a batch. The post says the accuracy gain offsets a small latency cost. Separate heads cover estimates, forecasts, durations, calculated dates and explicit arithmetic. Confidence scores, character spans and probabilities remain on the response. PDFs are new relative to the preview, which the authors say could read a photo of a receipt. A request can include up to four images or PDF pages, and line items can come back as rows.
A receipt set the company marked itself
The scores are internal, and the post says so. “There are no public extraction benchmarks yet,” and the comparisons “are self-reported.” On a set of fresh text fields, 96 requests, the preview was right on 295 of 336, Clairic on 335 and Clairic-flash on 328. On synthetic receipt fields in text and images, the preview was right on 369 of 576 and both new models on 576 of 576. That is the 207-field gap the prose summarizes as image-reading errors falling from 207 to zero on an internal benchmark covering text and images. On a mixed-input set, 36 requests, the preview was right on 207 of 226, Clairic on 221 and Clairic-flash on 213. Two other models appear in the same table, Liquid AI’s D1.2 pipelined with Apple Vision, and Scaledown’s Extract. On the fresh text fields those scored 304 and 259 of 336. The authors call the table directional and say they wanted to show that other companies are shipping extraction models too.
The latency numbers are from a self-hosted server, measuring a complete HTTP response, not from the Workers AI endpoint. On the 96 fresh text requests, Clairic’s median was 358 milliseconds and its 95th percentile 736. Clairic-flash’s were 155 and 328. On 48 fresh image requests, Clairic was 453 and 976, and the flash model 305 and 605. The post’s claim for that setup is a response in under a second. It does not publish a Workers AI latency for the same sets. The team name in the post is new as well. Reneau, Chen and Flansburg call themselves the Workers AI superintelligence team, and they name Clef and Clairic as the models behind that label. The open-weight release is the part a reader can download. The $0.24 price is the part that starts on 1 December.


