Key takeaways
- Cloudflare said on Nov. 18 that Clairic now runs as a 27 billion parameter model at its existing Workers AI endpoint, while the 9 billion parameter version moves to a new Clairic-flash endpoint.
- Clairic will cost $0.24 and Clairic-flash $0.10 per million input tokens, with both at $0.10 until Dec. 1, 2026, and Cloudflare does not charge for output tokens.
- Open weights for both models are on Hugging Face, and Cloudflare said its self-reported tests show perfect scores on synthetic receipts and tracked customer checks.
In this story 4 sections
Developers who already send requests to Cloudflare's Clairic endpoint are now getting a bigger model behind the same address, with no code changes. Cloudflare (NYSE: NET) said on Wednesday, Nov. 18, that Clairic, its data extraction model, has left preview as a 27 billion parameter model, and that a smaller 9 billion parameter version called Clairic-flash is joining it. Both now come with open weights on Hugging Face.
The update was announced in a Cloudflare blog post written by Alex Reneau, Michelle Chen and Kevin Flansburg, who describe themselves as the Workers AI superintelligence team. "Cloudflare has become a neolab," the post opens, saying the company now builds its own frontier models for customers and the open source community.
What Clairic does
Clairic is not a chatbot. It is built to pull specific fields out of messy input: a merchant name from a receipt photo, values from a support ticket, totals and line items from a PDF invoice. Cloudflare says answers still arrive with confidence scores, character spans and probabilities, so an app can check where each value came from.
The new version can read PDFs for the first time. Each request accepts a combined maximum of four pages or images, and Cloudflare says invoice lines are returned as grouped rows, so an item's name, count and price stay linked.
What changes for developers on Workers AI
- The endpoint gets an upgrade in place. The preview at `@cf/cloudflare/clairic` was actually the 9B model. That address now serves the 27B Clairic, and Cloudflare says existing apps will pick up the better extraction automatically.
- The 9B model moves. It now lives at `@cf/cloudflare/clairic-flash` for teams that care more about cost and speed.
- Pricing splits on Dec. 1. Clairic will cost $0.24 per million input tokens and Clairic-flash $0.10. Until Dec. 1, 2026, both are $0.10 as a promotion, which by our math is about 58% below Clairic's standard rate ($0.14 saved on $0.24).
- Output is free. Cloudflare bills only input tokens. "We eat the cost of the output token generation," the post says.
- The API schema stays the same as the preview, according to the company.
How Cloudflare says it performs
Cloudflare said it rebuilt the architecture after the preview. Text and number fields are now generated token by token in parallel slots instead of in one forward pass, and dedicated heads were trained for math, forecasts, time spans, date calculations and estimates.
On its own tests, the company reported these results against the preview model and two outside extraction tools, Liquid AI's D1.2 paired with Apple Vision OCR and Scaledown's Extract:
- Synthetic receipt fields: 576 of 576 correct for both new models, compared with 369 for the preview, 397 for D1.2 with OCR and 280 for Scaledown.
- Tracked customer assertions: 1,263 of 1,263 for both new models, up from 1,033 for the preview.
- Fresh text fields: 335 of 336 for Clairic and 328 for Clairic-flash.
- Image reading: errors on internal benchmarks fell from 207 to zero.
Cloudflare stresses that no public extraction benchmark exists yet and that these figures are self-reported, so readers should treat them as directional.
Speed is where the smaller model earns its place. On fresh text fields, Clairic-flash returned a median response in 155 milliseconds against 358 for Clairic, about 57% less time by our calculation. Cloudflare says both stay under one second in typical use, measured as a full HTTP response from a self-hosted server.
Why the open weights matter
Releasing the weights lets researchers and companies download, test and fine tune both models outside Cloudflare's cloud. The company said it saw many experiments and tuned versions after it opened its earlier Clef model, and it wants the same for Clairic. Its stated aim is to help customers move narrow jobs off expensive general purpose language models and onto smaller, cheaper ones built for a single task.
Smaller, cheaper models have been a steady theme this month, from Inkling 2's small weights on Hugging Face to Google's mid tier Gemini 4 Neon. More model releases are tracked on our AI news page.
Share this story
Spotted an error? Tell us


