Last updated: November 19, 2026
Cloudflare Clairic is now an open-weight extraction model family. On November 18, 2026, Cloudflare moved a redesigned 27B Clairic to @cf/cloudflare/clairic, added a 9B Clairic-flash, added PDF input, and published both models on Hugging Face. Each costs $0.10 per million input tokens until December 1.
Contents
- What is new in Clairic
- Clairic vs. Clairic-flash
- How to call Clairic on Workers AI
- Field types and response data
- How to run the open weights
- Pricing and billing
- Limits
- Workflows, limitations and alternatives
- Frequently asked questions
What is new in Clairic
Cloudflare announced the release in a blog post dated November 18, 2026. Cloudflare calls Clairic a System One model: a small, typed model built to turn unstructured input into structured values more cheaply than a large LLM.
The main changes since the preview are:
- A new architecture. Text and number fields now use parallel slot autoregressive generation instead of a single forward pass. Each field writes its own tokens and stops at its own end marker, and all fields advance together in one serving batch.
- Special heads for estimates, forecasts, durations, calculated dates and explicit arithmetic.
- PDF input. You can ask for a supplier, line items and a total in one request, even when they sit on different pages.
- Two sizes and open weights on Hugging Face for both models.
The typed API did not change. Cloudflare says confidence scores, character spans and probabilities still come back with every request, and the API schemas are unchanged from the preview endpoint. Apps already calling the preview endpoint get the new 27B model automatically as it rolls out in place.
Clairic vs. Clairic-flash
The table combines the Cloudflare blog, the Clairic model page, the Clairic-flash model page and the Hugging Face model cards. Latency figures are Cloudflare’s own self-hosted tests on its “fresh text fields” set.
| Item | Clairic | Clairic-flash |
|---|---|---|
| Workers AI model ID | @cf/cloudflare/clairic | @cf/cloudflare/clairic-flash |
| Size (per blog) | 27B parameters | 9B parameters |
| Base model (per model card) | Qwen3.8-27B | Qwen3.5-9B |
| Price until Dec. 1, 2026 | $0.10 per M input tokens | $0.10 per M input tokens |
| Price after promotion | $0.24 per M input tokens | $0.10 per M input tokens |
| Context window | 65,536 tokens | 65,536 tokens |
| Inputs | Text, images, PDFs | Text, images, PDFs |
| Docs status | Preview | Early research preview |
| Latency p50 / p95, fresh text fields | 358 / 736 ms | 155 / 328 ms |
| Open-weight license | Apache-2.0 | Apache-2.0 |
On Cloudflare’s self-reported evals, the new Clairic scored 335 of 336 on fresh text fields versus 295 for the preview, and Clairic-flash scored 328. Both scored 576 of 576 on synthetic receipt fields in text and images, against 369 for the preview. Cloudflare says to treat these as directional, since there are no public extraction benchmarks yet.
How to call Clairic on Workers AI
- Create a Cloudflare API token with Workers AI access and note your account ID, or add an AI binding to a Worker.
- Choose the model:
@cf/cloudflare/clairicfor best accuracy or@cf/cloudflare/clairic-flashfor lower latency and cost. - Put your source text in
state. Add up to four images inimagesor PDFs indocumentsas embedded data. Remote URLs are not accepted. - Define
fields: a map of field names to typed specs, for example{"total": {"type": "number", "decimals": 2, "unit": "USD"}}. Add shortinstructionswhere a field could be ambiguous. - Send a POST to
/client/v4/accounts/{ACCOUNT_ID}/ai/run/@cf/cloudflare/clairic, or callenv.AI.run()from a Worker. - Read
answers, checktruncated, and apply your own confidence thresholds before you write values to a database.
Set reference_date to an ISO date when documents contain dates without a year or relative terms such as “net 30”.
Field types and response data
According to the Clairic model card, field specs support text, number, date, time, datetime, duration, boolean, enum, object, list, compute, estimate and forecast. Fields are nullable by default. Setting source_only: true restricts an answer to literal printed numbers or dates, and knowledge: true allows answers from model knowledge for supported types. Estimates and forecasts require bounds, and forecasts also take a horizon and interval.
Each answer is keyed by field name. Generated answers carry valid, complete and confidence data plus a separately scored p_null. Text evidence that matches the source includes character spans. Estimates can return a median and quantiles when you set options.include_quantiles. The usage object reports state, image and schema tokens, plus the number of passes.
How to run the open weights
Both repositories, Cloudflare/clairic and Cloudflare/clairic-flash, include safetensors weights, typed output components, tokenizer and processor files, Python helpers and a requirements file.
- Use Python 3.11 or later and install the packages in
requirements.txt. Inference uses PyTorch and Transformers, with xgrammar for schema constraints. - Download the repository with
snapshot_download("Cloudflare/clairic")and add the path tosys.path. - Load the model with
load_release_model(path, device="cuda")fromclairic_model. - Call
systemone(model, processor, request)with the same state and fields format as the hosted API. Pass PIL images for image input. - Use
predict_asyncin async applications. The helper raises a queue-full error for simultaneous requests, so add your own queue for concurrency.
Pricing and billing
Cloudflare bills only input tokens and absorbs the cost of output generation. Images and PDF pages are converted to input tokens. The Clairic-flash page explains the math: each 32×32-pixel block is one token, plus two marker tokens per image, and small images are scaled up to about 256×256.
One detail matters for cost. Each text field gets its own pass, other fields share a pass, and every pass includes the full state and all images. A request with many text fields over a large scanned page can therefore bill the page several times. Check usage.passes while you tune schemas.
Limits
- Up to four images or rendered PDF pages in total per request.
- 4 MiB per image or PDF, 8 MiB combined, and a 16 MiB request body.
- PNG, JPEG or WebP images. Encrypted PDFs are rejected.
- Each prompt, state plus fields, must fit in 65,273 input tokens. The schema says oversized requests are rejected rather than truncated.
Workflows, limitations and alternatives
Good fits are receipts, invoices, tickets and leases with known fields. Route low-confidence or high-p_null answers to a human or a larger model instead of trusting every value.
The four-page limit is the main constraint. Split longer PDFs and merge results yourself. Both hosted models are still labeled preview in the docs, and Cloudflare’s terms say it may stop offering a model at any time. Clairic also is not a chat model; for open-ended reasoning over documents, see guides such as whether Claude can read PDFs.
For other typed-output models, compare our guides to Liquid AI d1.2 typed answers and pricing, the Gemini Decision 1 API and Perplexity Decider v1.3. For Cloudflare’s agent tooling, see the Cloudflare Kitesurf guide.
Frequently asked questions
What is Cloudflare Clairic?
Clairic is a Cloudflare extraction model. You send source text, images or PDF pages plus a schema of typed fields, and it returns structured answers with confidence scores, null probabilities and character spans. It can also compute values, estimate numbers and forecast series.
What is the difference between Clairic and Clairic-flash?
Clairic is the 27B model at @cf/cloudflare/clairic. Clairic-flash is the lighter 9B model at @cf/cloudflare/clairic-flash. Cloudflare says Clairic-flash is faster, while Clairic scored higher on its self-reported extraction evals.
How much does Clairic cost on Workers AI?
Until December 1, 2026, both models cost $0.10 per million input tokens. After that, Clairic returns to $0.24 per million input tokens and Clairic-flash stays at $0.10. Cloudflare does not charge for output tokens.
Are the Clairic weights open source?
Both models are published on Hugging Face as open weights. Their model cards state that they are released under the Apache-2.0 license, and they run locally with PyTorch and Transformers.
Can Clairic read PDFs?
Yes. The updated models accept embedded PDFs. Rendered PDF pages and images together are limited to four per request, each file can be up to 4 MiB, and encrypted PDFs are rejected.
Sources
- Cloudflare Blog: An updated Clairic and Clairic-flash are now open-weight (Nov. 18, 2026)
- Cloudflare Workers AI docs: clairic
- Cloudflare Workers AI docs: clairic-flash
- Hugging Face: Cloudflare/clairic model card
- Hugging Face: Cloudflare/clairic-flash model card
Image: Cloudflare logo, public domain, via Wikimedia Commons. Cloudflare is a trademark of Cloudflare, Inc.