Clairic is a model that extracts specific values from unstructured text and images. You describe the values you want, such as an order number, a total, or a date. Clairic returns each one in the type you asked for, along with a confidence score your code can check.
Cloudflare calls Clairic a System One extraction model. It reads the input in one pass and returns a value for each field without generating text. Michelle Chen announced it on X, and it is available as a preview on Workers AI with the model ID @cf/cloudflare/clairic.
The best-known System One model is Jev from TypeSafe, and the two have different purposes. Jev chooses an answer from options you define, such as approve, review, or reject. Clairic returns values you don't know in advance, such as a total or a date. They often work together, and the Clairic and Jev section below shows how.
Why use Clairic
You can already ask a large language model to return JSON. It usually works, but you still have to parse the reply, check its shape, and guess whether each value is right.
Clairic removes most of that work. Every answer comes back in the type you declared, and a missing value comes back as null with a probability instead of an invented value.
Use Clairic when
you need a few specific values, not an explanation, you know the type of each value in advance, you want a score that shows which answers a person should review.
What Clairic can return
Clairic supports four tasks.
Supported field types include text, number, date, time, datetime, duration, list, estimate, and forecast. Text answers can include the exact position in the source, so you can highlight where they came from. The model reference lists the full request and response format.
Support intake. Extract the order ID, delivery date, amount paid, and damaged parts from a customer email. Send low-confidence fields to a person before opening a claim.
Invoices and receipts. Read totals, dates, names, and line items from text or photos.
Quick estimates. Estimate a repair cost or payment time from examples in the input. The range shows how uncertain the estimate is.
Short forecasts. Predict the next few values in a pasted series, such as daily ticket volume, when you do not need a dedicated forecasting system.
Call it from code
Here is the support email from the playground.
In a Cloudflare Worker with a Workers AI binding, pass the message as state and list the values you want in fields.
The expected response has this shape. Confidence values can vary, but the structure stays the same.
Two values are computed, not copied. Clairic adds the item cost and tax to get 701.10, and counts the days between the two dates to get 11. The tracking number is not in the message, so it comes back as null.
For exact math in production, have Clairic extract the raw prices and dates and do the calculation in your own code.
Clairic and Jev
Clairic is one of three model types you might use together.
A useful pattern pairs Clairic with a decision model such as Jev 1.14. Clairic extracts the facts from a claim, and Jev chooses the outcome your policy allows.
Use Clairic when you know the type of the answer but not the value. Use Jev when you already know the allowed answers.
When not to use Clairic
Open-ended writing or reasoning. Use a generative model.
Picking from fixed options. Use a decision model such as Jev.
Anything code can do exactly. Arithmetic, permissions, IDs, and irreversible actions belong in deterministic code.
Production forecasting. Clairic only sees the series you give it and cannot know about events missing from the data.
Limitations and risks
Clairic is a preview, so its behavior and availability may change. Cloudflare reports that one- and two-step calculations are reliable, but longer chains and totals across many rows are still hard.
A typed answer can still be wrong. Confidence scores and 80% ranges are useful signals, but they are not equally well calibrated on every task. Clairic accepts up to four PNG, JPEG, or WebP images per request, and values read from hard-to-read photos can be wrong. Check high-impact values against the source, your business rules, or a person.
Next steps
Pick one repeated extraction task with a few clearly named fields.
Test it on real examples, including missing values and noisy images.
Set review rules based on what a wrong answer would cost.
Keep exact checks in code and pass only verified values to later steps.
Read Cloudflare's Clairic announcement, the extraction model guide, and the model reference before using the preview in production.
Order K-271 arrived on 2026-10-02 with a cracked screen and bent frame. The item cost $649.00 and tax was $52.10. I am writing on 2026-10-13. Please repair it or refund me.
