Cloudflare's Clairic pulls typed values from text and photos in one pass, and claims 3.2x over an LLM

What happened

Cloudflare released Clairic, an extraction model, in preview on Workers AI as @cf/cloudflare/clairic, per a Cloudflare post by Michelle Chen dated 19 October 2026. An extraction model reads text and images and returns specific values, such as an order number, a date or a price. It does not write sentences. Cloudflare says it processes the input in a single pass with no output tokens and no reasoning step, and calls it a new class of System One-style models, a term from TypeSafe, whose Jev model the post credits (we covered Jev 1.14). Cloudflare says it intends to release open weights "soon", with no date and no licence stated. Architecture, size and training data are not disclosed.

Model

@cf/cloudflare/clairic, preview on Workers AI

Input

Text plus up to four images

Output

Typed fields, each with confidence and a null option

Pricing

Charged for input only; no per-token rate in the post

Weights

Planned open, no date

As of

20 Oct 2026

How you use it

A request holds a state (the text and optional images) and a set of named fields. Each field has a type and optional plain-language instructions. Types are text (returns the span and its position), number, date, time, datetime, duration, list, estimate and forecast.

Scalar answers carry p_null (the chance the value is absent) and confidence. An estimate returns a median with an 80% interval (10th to 90th percentile) on a range you set. A forecast returns one estimate per future step. The model returns null when the input does not contain the value. If many fields do not fit one pass, the server splits them across parallel passes and merges the answers.

What it showed

All numbers below are Cloudflare's own and have no independent confirmation. The post gives no benchmark suite, sample size or dataset beyond its examples.

damaged-product claim pipeline, Clairic plus Clef vs one LLM (GLM-5.3)

19.3 s vs 61.7 s

cost per million claims, 49% lower

$6,032 vs $11,740

Clairic extraction step

0.46 s

TSA daily forecasts inside the 80% interval

17 of 21

The claim demo chains Clairic to Clef, Cloudflare's decision model, which picks one answer from a fixed list. In both pipelines the facts and the decision came out right every run. Clairic's repair estimate was identical each time, while the LLM's varied ($250, $250, $220). Cloudflare says Gemma 4 and Qwen could not derive values such as the repair cost in the all-LLM version.

Other examples in the post: a support email returned the order ID, delivery date, amount paid and days since delivery, with the tracking number correctly null. A receipt photo returned order number, quantity, prices and tip. A home-sale estimate gave a median of 415,000 with a range of 370,000 to 470,000. A forecast of daily TSA passenger counts averaged 3.3% error over 21 days, and three of its four misses fell on the Columbus Day weekend.

Where it could be used

The post's own examples point to jobs where the same fields are pulled again and again from messy input:

• Support and claims intake: read an email or a photo, pull the order ID, dates and amounts, then route the case.
• Receipts and invoices: photo to line items, totals and tax.
• Triage by confidence: send anything with a high p_null or low confidence to a person, as Cloudflare suggests.
• Rough estimates: a price or lead-time range with an interval, when the input holds the evidence.
• Pipelines with a decision step: extraction first, then a fixed-answer model or plain code.

These match the post's examples. Uses beyond them, such as form processing or log parsing, are reasonable guesses and not tested claims.

What to watch

Watch for the open-weights release and its licence, a per-input price, and tests by people outside Cloudflare on their own documents. Everything above rests on one vendor post.

Sources

1. Cloudflare: Clairic extraction model: model, request format, examples, demo figures, pricing basis, limits, open-weights plan