Sulat.com / Writing on AIThursday, November 19, 2026
Artificial Intelligence/19 November 2026

Cloudflare releases Clairic and Clairic-flash as open weights, and rebuilds Clairic so each field writes its own tokens

Cloudflare rebuilt Clairic, added a 9B Clairic-flash and released both as Apache-2.0 open weights. Both cost US$0.10 per million input tokens until 1 December, and the benchmarks are Cloudflare's own.

Glossy 3D Cloudflare logo beside three glass documents with highlighted fields pulled out as typed values

At a glance

  • Cloudflare says it has rebuilt Clairic (now 27B) and released it and a new 9B Clairic-flash as Apache-2.0 open weights on Hugging Face, also hosted on Workers AI.
  • For ordinary text and number fields, each field now writes its own tokens until an end marker, where the October preview said Clairic generated no tokens. Output tokens are still free.
  • Clairic's docs bill the input once per pass: each text field gets its own pass, and the docs example turns 61 tokens into 655 input tokens over three passes.
  • Cloudflare's benchmarks are self-reported. Both new models score 576 of 576 on its receipt test, where the preview scored 369, and latency was measured on a self-hosted server.
  • Both models cost US$0.10 per million input tokens until 1 December; after that Clairic is US$0.24 and Clairic-flash stays at US$0.10.

Cloudflare says it has rebuilt Clairic, its extraction model, and released it and a new 9-billion-parameter sibling, Clairic-flash, as open weights under the Apache-2.0 licence on Hugging Face. Both are also hosted on Workers AI, Cloudflare’s service for running models on its network. The announcement is a blog post by Alex Reneau and Michelle Chen, published at 22:00 UTC on 18 November, which is 11am on 19 November in New Zealand. An extraction model reads text, images or PDF pages and returns typed values, such as a supplier name, a total or a due date.

Two models, two endpoints

Clairic is now the 27-billion-parameter model and Clairic-flash is the 9-billion-parameter one. The preview, announced on 19 October, ran the 9B model at US$0.10 per million input tokens. That model has moved to the flash endpoint, and Cloudflare says anything wired to the old endpoint gets the new Clairic rolled out in place.

ModelSizeWorkers AI IDInput price per million tokensBase model on the Hugging Face card
Clairic27B@cf/cloudflare/clairicUS$0.24 (US$0.10 until 1 December)Qwen3.8-27B
Clairic-flash9B@cf/cloudflare/clairic-flashUS$0.10Qwen3.5-9B
Sources: Cloudflare’s blog post and the two Hugging Face model cards. Output tokens aren’t billed.

The team behind both is what Cloudflare now calls the Workers AI superintelligence team, which also made Clef, its decision model released on 1 October. The post opens with the line “Cloudflare has become a neolab.” Clairic and Clef belong to what Cloudflare and TypeSafe call System One models: small models that return typed answers in a single request instead of free text. TypeSafe launched the category with its Jev model on 15 September.

The rebuild: each field writes its own answer

The preview’s pitch was speed through a single forward pass. Cloudflare’s October post said Clairic and Clef “answer in a single pass, without reasoning or generating tokens”, and that because Clairic generates no tokens it can’t produce a mistyped answer or “hallucinate” in the same way as an LLM.

The new post changes that for most fields. For ordinary text and number fields, Cloudflare now uses what it calls parallel slot autoregressive generation. Each field uses the tokens it has already written to choose its next token, then stops at its own end marker, and all the fields advance together in the serving batch. Cloudflare says this raised accuracy for text and images and that the gain offsets “the small latency cost”. For computation, estimation and forecasting, it trained separate heads that handle estimates, forecasts, durations, calculated dates and explicit arithmetic.

Preview answers all fields in one pass; the new Clairic lets each field write its own tokens until an end marker
The preview answered every field at once; the new version lets each field write its own tokens.

The post says the model “does technically generate output tokens” but keeps the System One pricing rule of charging for input only. The Hugging Face cards add that schema constraints come from xgrammar, a Python library, applied through a Transformers logits processor. The new post doesn’t repeat the preview’s claim that Clairic can’t hallucinate the way an LLM does.

The typed API is unchanged. Every answer still comes back with a confidence score, character spans and probabilities, and Cloudflare says the API schemas are unchanged from the preview endpoint.

Billing counts the input once per pass

Clairic’s docs page explains a cost detail the blog post doesn’t. Clairic answers each text field in its own pass, and the other fields share a pass. Fields that don’t fit beside the input are split across more passes. Every pass includes the full input and all images, so those tokens are billed once per pass, and usage.passes reports the count.

Here’s the docs example. An invoice of 61 tokens, six fields (two of them text) and 472 schema tokens came to 655 input tokens over three passes, which is 61 × 3 + 472.

Three passes each carry the 61-token state; 61 x 3 plus 472 schema tokens makes 655 input tokens
Cloudflare’s docs example bills the 61-token input three times, once per pass.

At scale that adds up. A 20,000-token document with four text fields and a number field would run five passes and bill about 100,000 input tokens before the schema, which is US$0.01 at US$0.10 per million and US$0.024 at US$0.24. That’s my arithmetic from the docs’ rules, not a figure Cloudflare published. Images cost one token per 32×32-pixel block plus two marker tokens, so a 1,024×768 receipt photo is 770 tokens on every pass. Each request takes up to four images, and a prompt has to fit in 65,273 input tokens.

The benchmarks are Cloudflare’s own

Cloudflare says there are no public extraction benchmarks yet, so it ran internal evals and asks readers to take them as “directional guidance since they are self-reported”. The two rivals are Liquid AI’s D1.2, which Cloudflare paired with Apple Vision for text recognition, and Scaledown’s Extract. The table shows Cloudflare’s counts with the percentages worked out.

EvalClairic-previewClairicClairic-flashD1.2 + OCRScaleDown
Fresh text fields (96 requests)295/336 (87.8%)335/336 (99.7%)328/336 (97.6%)304/336 (90.5%)259/336 (77.1%)
Mixed-input fields (36 requests)207/226 (91.6%)221/226 (97.8%)213/226 (94.2%)116/226 (51.3%)156/226 (69.0%)
Tracked customer assertions (139 request shapes)1,033/1,263 (81.8%)1,263/1,2631,263/1,263973/1,263 (77.0%)653/1,263 (51.7%)
Synthetic receipt fields, text and images369/576 (64.1%)576/576576/576397/576 (68.9%)280/576 (48.6%)
Cloudflare’s self-reported results. Percentages are my arithmetic. The mixed-input row is labelled “Internal eval (mchen)” in the post.

The post says image reading accuracy improved “from 207 errors to 0”. That 207 matches the preview’s misses on the receipt row, 576 minus 369. The post’s example is a synthetic coffee shop receipt where the preview returned null for the merchant name and Clairic returned “Coffee House”, with all five fields right. The post doesn’t publish the test sets or the prompts.

Bars of correct receipt fields out of 576: Clairic 576, Clairic-flash 576, D1.2 plus OCR 397, preview 369, ScaleDown 280
On Cloudflare’s own receipt test, the preview missed 207 fields and the new models missed none.

PDF reading is new. Clairic could read a photo of a receipt in the preview. Cloudflare says the family now reads PDFs too, so one request can ask for the supplier, line items and total even when they sit on different pages, with line items returned as rows that keep description, quantity and price together. The limit is four images or PDF pages per request.

Latency: under a second, measured off-platform

Cloudflare measured the full HTTP response, including every field in the request, from a self-hosted server. It wasn’t measuring Workers AI. Its slowest result is a 976 ms p95 for Clairic on image fields. Clairic-flash’s median is between 1.5 and 2.3 times faster than Clairic’s across the six sets, by my arithmetic.

Eval setRequests per modelClairic p50 / p95Clairic-flash p50 / p95
Fresh text fields96358 / 736 ms155 / 328 ms
Fresh image fields48453 / 976 ms305 / 605 ms
Commercial leases180691 / 803 ms379 / 519 ms
Loan pricing600395 / 407 ms188 / 192 ms
Traffic forecasting114532 / 585 ms243 / 261 ms
Expense audit, development60632 / 742 ms393 / 512 ms
Latency figures from Cloudflare’s post, measured on a self-hosted server.

Pricing: a promotion until 1 December

Until 1 December 2026, Clairic and Clairic-flash both cost US$0.10 per million input tokens. On 1 December Clairic goes to US$0.24 and Clairic-flash stays at US$0.10. The docs page calls the promotion 58% off. Output tokens are free on both, which Cloudflare says follows System One pricing principles.

For comparison, Liquid AI lists D1.2 at US$0.02 and D1.2-ultra at US$0.042 per million input tokens in its 16 November announcement, and TypeSafe lists Jev at US$0.042. OpenRouter’s catalogue, dated 6 November, lists Scaledown Extract at US$0.04 and Clairic at US$0.10. Those are list prices for models with different jobs, and Clairic’s per-pass billing means the price per request depends on how many text fields you ask for.

Running the weights yourself

Both repositories carry the Apache-2.0 licence and ship as BF16 safetensors. Hugging Face lists Clairic at 28B parameters and Clairic-flash at 10B, which comes to about 56 GB and 20 GB for the weights alone at two bytes per parameter. That’s my arithmetic, and it leaves out the memory the model needs while running.

The cards say to use Python 3.11 or later with PyTorch, Transformers and xgrammar. Each repository includes a clairic_model.py helper with a systemone function and a clair_head.safetensors file for the typed outputs. The local helper takes embedded inputs of up to 4 MiB each and 8 MiB combined, within the four-image limit, and it doesn’t accept remote media URLs.

On Workers AI, a request looks like this, trimmed from the docs example:

const response = await env.AI.run("@cf/cloudflare/clairic", {
  model: "clairic",
  state: "Invoice INV-2291 from Northwind Supply, issued 2026-09-02, net 30. 4 pallets at $250.00 each, shipping $80.00.",
  fields: {
    invoice_number: { type: "text" },
    due_date: { type: "date", instructions: "Issue date plus the net payment days" },
    total: { type: "number", decimals: 2, unit: "USD", instructions: "Amount due including shipping" },
    po_number: { type: "text", instructions: "Purchase order number" }
  }
});

In the docs example the invoice has no purchase order number, so po_number comes back null with a p_null of 1. Null is how the model says the input doesn’t contain the value.

What the docs still say

Clairic’s docs page still calls the model “in preview”, and the Clairic-flash page calls it an “early research preview”, although the blog post says Clairic graduates from preview. The flash page lists US$0.10 and doesn’t carry the promotion banner. The post doesn’t describe how the 27B model was trained or on what data.


Discover more from Sulat AI & Tech News

Subscribe to get the latest posts sent to your email.