Cloudflare has released two open-weight Clairic models that turn receipts, PDFs, and text into structured data.
The models are a developer tool, not a consumer app. Extraction means turning receipts, PDFs, or text into named fields such as supplier, line items, and total. The smaller Clairic-flash costs $0.10 per million input tokens on Workers AI, and Unite.AI reports developers can also run it on their own hardware.

Cloudflare’s 18 November release pairs the updated, 27-billion-parameter Clairic with the new 9-billion-parameter Clairic-flash. Both are hosted on Workers AI, and the full weights are on Hugging Face.
What’s changed since the October preview
The October preview was a hosted endpoint running the 9B model. The updated Clairic can now read PDFs and images, pulling the supplier, line items, and total in one request even when they appear on different pages. Each request can include up to four images or PDF pages in total.

Cloudflare redesigned how the models fill ordinary text and number fields. Instead of producing them in a single pass, the models build each field step by step while working on multiple fields in parallel. Separate components handle estimates, forecasts, durations, calculated dates, and arithmetic.
The typed API still returns confidence scores, character spans, and probabilities. Cloudflare says customers asked for use cases from extracting data from receipts to reading values from support tickets.
Clairic vs Clairic-flash: accuracy and speed
There aren’t any public extraction benchmarks yet, so these are the company’s own internal evaluations, not independent results. Image-reading errors fell from 207 to zero in those tests.
| Model or system | Fresh text fields (96 requests) | Tracked customer assertions (139 request shapes) | Synthetic receipt fields |
|---|---|---|---|
| Clairic preview | 295/336 | 1,033/1,263 | 369/576 |
| Clairic (27B) | 335/336 | 1,263/1,263 | 576/576 |
| Clairic-flash (9B) | 328/336 | 1,263/1,263 | 576/576 |
| D1.2 + OCR | 304/336 | 973/1,263 | 397/576 |
| ScaleDown | 259/336 | 653/1,263 | 280/576 |
Clairic-flash matched the larger model on the customer assertions and synthetic receipts, but got seven fewer fresh text fields right. It was faster in all six latency tests. The timings below are Cloudflare’s own measurements from its self-hosted server, not independent benchmarks.
| Workload | Clairic median / 95th percentile | Clairic-flash median / 95th percentile |
|---|---|---|
| Fresh text fields | 358 / 736 ms | 155 / 328 ms |
| Fresh image fields | 453 / 976 ms | 305 / 605 ms |
| Commercial leases | 691 / 803 ms | 379 / 519 ms |
| Loan pricing | 395 / 407 ms | 188 / 192 ms |
| Traffic forecasting | 532 / 585 ms | 243 / 261 ms |
| Expense audit | 632 / 742 ms | 393 / 512 ms |
Workers AI pricing and how to run Clairic locally
| Model | Workers AI endpoint | Standard price per million input tokens (USD) |
|---|---|---|
| Clairic (27B) | @cf/cloudflare/clairic | $0.24 |
| Clairic-flash (9B) | @cf/cloudflare/clairic-flash | $0.10 |
Both models cost $0.10 per million input tokens until 1 December 2026, after which the 27B model returns to $0.24. Workers AI doesn’t charge for output tokens.
Existing apps on the preview Clairic endpoint automatically move to the updated model without an API schema change. The 9B option now has its own Clairic-flash endpoint and keeps the preview’s $0.10 rate.
Unite.AI says developers can run the models locally with PyTorch and Transformers without using Workers AI. Both Hugging Face repositories use the Apache-2.0 licence and include weights, typed-output components, tokeniser and processor files, and Python helpers.
Cloudflare also introduced its Workers AI Superintelligence research team, which built Clairic and Clef. The company bought Deno in October, and Tbreak covered Deno’s programming model.
Until 1 December, both models cost $0.10 per million input tokens. After that, the 27B model’s $0.24 rate applies, 2.4 times the Clairic-flash price.


















