Cloudflare Clairic: what it is and what it costs

Cloudflare Clairic is a new model on Workers AI that pulls typed values out of messy text and images, things like amounts, dates, durations, estimates and forecasts, instead of writing a reply. It is in preview as @cf/cloudflare/clairic, and Cloudflare's docs list it at $0.10 per million input tokens with no output pricing, because it doesn't generate output tokens at all.

Cloudflare announced it on Monday. The @Cloudflare post calls it "a System One extraction model," built for "extracting, computing, estimating, and forecasting data from unstructured inputs." Twelve minutes earlier, @michellechen, who wrote the launch blog post, pitched it as "a brand new model paradigm." Cloudflare's @kevin_flansburg explained the technical difference: Clairic is "based on regression instead of classification," so it can estimate a number without you listing every possible answer up front. He also repeated that open weights are coming.

What "System One" means here

Cloudflare's earlier model in this family, Clef, is a decision model: you give it a question and a set of options, and it returns odds for each one. That is the same idea Perplexity shipped a few weeks ago, which we covered in Perplexity Decisions API: what it is and what it costs. Clairic covers the other half. Clef answers "which one," and Clairic answers "what" and "how much."

It works in one parallel pass. There is no chain of reasoning and no token-by-token writing, so in Cloudflare's demo the extraction step took 0.46 seconds. Every answer comes back with a confidence score and a probability that the value isn't in the input at all. If a receipt has no discount line, you get null instead of a made-up number.

How to try Clairic

You call it from a Cloudflare Worker or through the REST API. You send a state (the text, plus up to four images) and a set of named fields, each with a type and optional plain-English instructions. Here is the shape, using the field types from Cloudflare's docs on a service-request email:

const { answers } = await env.AI.run("@cf/cloudflare/clairic", {
  state: email,
  fields: {
    customer_name: { type: "text" },
    preferred_date: { type: "date", instructions: "Day the customer asked for" },
    quoted_total: { type: "number", decimals: 2, unit: "USD" },
    job_value: { type: "estimate", unit: "USD", min: 100, max: 50000, scale: "log", step: 50 }
  }
});

The field types available in the preview:

You can ask for as many fields as you want. If they don't fit in one pass, Cloudflare splits them across parallel passes that each see the whole input and merges the results.

What Clairic costs, and is it free?

The model page lists $0.10 per million input tokens and a 65,536-token context window. Workers AI accounts also get a free daily allocation of 10,000 Neurons that every model draws from. Cloudflare doesn't say how many Clairic calls that covers, so test with your own documents before you count on it.

The bigger cost claim is about the full pipeline. In Cloudflare's own test on customer claim emails, Clairic extracted, Clef decided and GLM-5.3 wrote the reply. That setup ran about 3.2 times faster and cost 49% less than using GLM-5.3 for every step. Keep in mind that Cloudflare ran that benchmark on its own task.

Where it falls short

Cloudflare lists the weak spots itself. One or two steps of math are reliable, but long calculations and totals across many table rows are not, and the company suggests doing that math in code. Estimates need evidence in the input. In its 21-day TSA forecasting test, the average error was 3.3%, but 4 of 21 actual counts fell outside the 80% range, and it missed holiday spikes. The endpoint is an early checkpoint. Answers always come back well formed, right or wrong, so you lean on the confidence scores to decide what runs on its own and what a person reviews.

What this means for a small business

Most of the paperwork drag in a service business is extraction: reading an email or a photo of a receipt and typing six values into another system. A no-heat call that shows up as "furnace not working" and a phone number is the kind of thin intake we wrote about in why South Elgin no-heat calls arrive without the details. A model that fills those fields in half a second, and tells you when a field is missing, is a cheap first filter in front of a person.

If you want to know which of your forms, inboxes or invoices are worth putting through a model like this, New Face Design's free process audit maps where your hours go and picks the few steps worth automating first.