Cloudflare has released an updated version of its Clairic model and introduced Clairic-flash, a smaller variant designed for faster, lower-cost extraction workloads. Both models are available through the company’s Workers AI platform, with their full weights published on Hugging Face.
The release follows a redesign of Clairic after its preview. Cloudflare says the revised model is better at extracting information from unstructured text, images and PDFs, including tables, lists and dates. It is aimed at tasks such as reading receipt details or retrieving information from support tickets, rather than serving as a general-purpose language model.
Clairic-flash has 9 billion parameters. The standard Clairic model has 27 billion parameters, according to Cloudflare’s pricing and endpoint details. Developers can access them as @cf/cloudflare/clairic and @cf/cloudflare/clairic-flash, respectively. Cloudflare says the API schema remains unchanged from the preview endpoint.
Redesigned generation for extraction
Cloudflare says the new Clairic uses parallel slot autoregressive generation for ordinary text and number fields. Rather than producing all extracted values in a single forward pass, each field uses its previously generated tokens to produce its next token and stops at its own end marker. The fields advance together in the serving batch.
The company says this approach improved extraction accuracy while adding a small latency cost. For computation, estimation and extraction tasks, Cloudflare also trained specialized model heads to handle estimates, forecasts, durations, calculated dates and explicit arithmetic. The typed API continues to return confidence scores, character spans and probabilities, features intended to help developers assess and use the extracted results.
Cloudflare reported results from several internal evaluations, while noting that public extraction benchmarks are not yet available. On a set of 96 requests for fresh text fields, Clairic answered 335 of 336 fields correctly, compared with 328 for Clairic-flash and 295 for Clairic-preview. On an internal mixed-input evaluation, the new Clairic scored 221 of 226, while Clairic-flash scored 213 and the preview model scored 207.
For synthetic receipt fields drawn from text and images, both new models returned all 576 fields correctly in Cloudflare’s evaluation; the preview model returned 369 correctly. Cloudflare also said its newer Clairic models reduced errors from 207 to zero on an internal image-reading benchmark. These figures are company-reported results and should be treated as directional rather than independent comparisons. Cloudflare included results for other extraction products in some tests, but said the evaluations were self-reported and not public benchmarks.
PDF support and reported latency
The models can process up to four images or PDF pages in a request, according to Cloudflare. The company says users can extract information such as a supplier, line items and a total even when those details appear across multiple pages. Line items can be returned as rows that keep descriptions, quantities and prices together.
Cloudflare also published latency measurements from tests that measured the complete HTTP response from a self-hosted server. In its fresh text-fields set, Clairic’s reported median and 95th-percentile response times were 358 milliseconds and 736 milliseconds; Clairic-flash recorded 155 and 328 milliseconds. For fresh image fields, the respective figures were 453 and 976 milliseconds for Clairic, and 305 and 605 milliseconds for Clairic-flash. These results reflect Cloudflare’s test setup and workloads, rather than a universal performance guarantee.
Pricing and availability
Cloudflare is charging for input tokens only, despite the models also generating output tokens. The company said Clairic will cost $0.24 per million input tokens and Clairic-flash $0.10 per million input tokens after its promotional period. Until December 1, 2026, both models are priced at $0.10 per million input tokens. Customers using the preview Clairic endpoint are being moved to the updated model in place, Cloudflare said.
The releases are part of Cloudflare’s Workers AI model lineup. The company also identified the Workers AI superintelligence team as the research group behind Clairic and Clef, and said it intends to develop additional models for Cloudflare customers and the open-source community. Making Clairic’s weights available gives developers and researchers the option to experiment with the models beyond the hosted Workers AI endpoints.