Cloudflare Open-Sources Clairic: Receipt Accuracy Jumps From 64% to 100%
Cloudflare has open-sourced the weights of Clairic (27B parameters) and Clairic-flash (9B), which pull the fields you need from a receipt, invoice or PDF: vendor, line items, total. This kind of routine work often goes to an expensive large LLM. On an internal test of 576 fields from synthetic receipts, the new Clairic filled in 100% correctly, against 64% for the preview version, and a response arrives in under a second. If you process a stream of similar documents, compare it with your current LLM on real files.
AI-processed from Cloudflare AI Blog; edited by Hamidun News
Every product has a boring task: pull the vendor, line items and total out of an email, an invoice or a photo of a receipt. It usually goes to a large language model, which is expensive and not especially fast. Cloudflare proposes a different approach: small models that do nothing but extract fields.
The company has updated Clairic (27B parameters), released a lighter Clairic-flash (9B), and opened the weights of both on Hugging Face. On an internal test with 576 fields from synthetic receipts, the new Clairic filled in 100% of fields correctly, while the preview version managed 64% (369 of 576). For developers, this means you can test parsing of receipts, requests and invoices on a specialized model instead of a general-purpose LLM.
What the models do and what has changed
You describe the fields you need, send text, an image or PDF pages, and get a typed response. For each field you receive a confidence score, the character position in the source text, and probabilities. This interface is unchanged from the preview, so you do not need to rewrite your integration.
The main addition is PDF reading. You can request the vendor, line items and total in a single call, even if they sit on different pages. Up to four images or pages are allowed. Table rows come back with the description, quantity and price kept together.
Generation has changed too. Previously, all fields were filled in a single pass. Now ordinary text and numeric fields are generated in parallel, but each field builds on the tokens it has already written and stops at its own end marker. The fields go through as one batch, so latency barely increases. Separate "heads" are trained for scores, forecasts, deadlines, computed dates and arithmetic. The 9B-parameter model from the preview has moved to the address @cf/cloudflare/clairic-flash, and @cf/cloudflare/clairic now serves the 27B version.
The numbers: accuracy and speed
In a test on fresh text fields (96 requests, 336 fields), Clairic answered correctly in 99.7% of cases (335 of 336), Clairic-flash in 97.6% (328), and the preview in 87.8% (295). The competitors: D1.2 from Liquid AI paired with Apple Vision for text recognition scored 90.5% (304), and Extract from Scaledown scored 77.1% (259). On tracked customer checks (1,263 items), both new models scored 100% against 81.8% for the preview (1,033). In a mixed text-and-image test (226 fields), Clairic got 97.8% (221), flash 94.2% (213), and the preview 91.6% (207).
Speed was measured on Cloudflare's own server rather than through Workers AI, and it covers the full HTTP response. On fresh text fields, Clairic's median is 358 ms and its 95th percentile (the latency within which 95% of requests are answered) is 736 ms; for flash the figures are 155 and 328 ms. With images, Clairic responds in 453 ms at the median and 976 ms at the 95th percentile, and flash in 305 and 605 ms. Credit calculation takes 395 ms for Clairic and 188 ms for flash, and traffic forecasting takes 532 and 243 ms.
There are caveats. There are no public benchmarks for data extraction yet, the tests are internal, and Cloudflare itself calls them a rough guide. The perfect 100% was achieved on synthetic receipts, which are easier than real crumpled photos. The blog post also contains an inconsistency: it states that image-reading errors fell from 207 to 0, but the 207 in the table is the number of correct fields for the preview, not errors.
Pricing and what to do
Clairic costs $0.24 per million input tokens, and Clairic-flash costs $0.10. Output tokens are not billed: Cloudflare absorbs their cost. Until December 1, 2026, both models are sold at $0.10, after which Clairic returns to $0.24. If you were using the preview endpoint, it will switch to the new model automatically.
The practical takeaway is simple. If you have a stream of similar documents, such as receipts, support tickets, lease agreements or invoices, run Clairic-flash and Clairic on your own files and compare accuracy and cost with your current LLM. The open weights let you deploy the model yourself or fine-tune it for your own formats. For one-off and poorly defined tasks, a general-purpose LLM is still more convenient.
Alongside the release, Cloudflare introduced the Workers AI Superintelligence team, which builds the company's own models, including the previously released Clef.
- Nov 18, 2026Willison Releases Three External Sandboxes So Agent-Run Code Stays Off Your Laptop
- Nov 18, 2026Dreamina Launches Motion Studio: Motion Graphics Written as Code From a Prompt
- Nov 18, 202612 Design Services Keep AI-Built Sites From Looking Like Slop: Free or From $6 a Month
- Nov 18, 2026sd.cpp instead of Forge Neo: nine models on a 16 GB GPU draw in 5–68 seconds
Key Questions Answered
What the models do and what has changed?
You describe the fields you need, send text, an image or PDF pages, and get a typed response. For each field you receive a confidence score, the character position in the source text, and probabilities. This interface is unchanged from the preview, so you do not need to rewrite your integration. The main addition is PDF reading. You can request the vendor, line items and total in a single call, even if they sit on different pages. Up to four images or pages are allowed.
What are the key details on The numbers: accuracy and speed?
In a test on fresh text fields (96 requests, 336 fields), Clairic answered correctly in 99.7% of cases (335 of 336), Clairic-flash in 97.6% (328), and the preview in 87.8% (295). The competitors: D1.2 from Liquid AI paired with Apple Vision for text recognition scored 90.5% (304), and Extract from Scaledown scored 77.1% (259). On tracked customer checks (1,263 items), both new models scored 100% against 81.8% for the preview (1,033).
What are the key details on Pricing and what to do?
Clairic costs $0.24 per million input tokens, and Clairic-flash costs $0.10. Output tokens are not billed: Cloudflare absorbs their cost. Until December 1, 2026, both models are sold at $0.10, after which Clairic returns to $0.24. If you were using the preview endpoint, it will switch to the new model automatically. The practical takeaway is simple.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.