JA EN
shiichan

Clairic, an Extraction Model That Answers in Typed Values, Lands on Workers AI!

Cloudflare Changelog ★★★★★ View Markdown
#cloudflare#ai#serverless

Hello, it's Shii-chan! A new kind of model has arrived on Workers AI today.

Cloudflare Changelog developers.cloudflare.com

What was announced?

According to the Cloudflare Changelog, @cf/cloudflare/clairic is now available on Workers AI as a preview. Clairic is the first "extraction model", a System One model that answers with values instead of text. Cloudflare also intends to release it with open weights soon.

Why it matters

Cloudflare explains that decision models such as Clef choose among options you define up front. But many structured-output tasks have answers you cannot list in advance, like an invoice total, a delivery date, the number of days a shipment is late, or a repair cost. Clairic handles those.

What changes

You send a state (the input) and a schema of typed fields, and you get back one typed value per field. A value is read from the input, computed from it, or predicted beyond it. That could cut down the work of asking an LLM for JSON and then validating it.

Dive Deep

Clairic does not generate tokens. It reads your input in one pass and decodes each answer under the rules of its type, so every answer is a valid value of that type, or null when the input does not determine it. Fields that do not fit in one pass are split across parallel passes, so there is no upper limit on fields.

The field types:

  • text: an exact span of the input, with character offsets
  • number: a number read or worked out from the input (honors integer and decimals)
  • date / time / datetime: real calendar dates and valid times in ISO form
  • duration: a length of time in the unit you choose
  • list: every matching text, number, or date, in input order
  • estimate: a distribution inside your min and max, on your step grid, on a linear or log scale
  • forecast: one estimate per future step of a series (up to 64 steps)

Main request fields:

  • state: the text to read (up to four images can be attached)
  • fields: a type per field name, with settings such as instructions, decimals, and unit
  • reference_date: a reference date for date-dependent fields
  • options.include_quantiles: returns all 99 percentiles for estimates and forecasts

Scalar answers include p_null (the probability the input does not determine the value) and confidence. Estimates and forecasts include a median and an 80% interval (p10 to p90). You can use these to decide what to automate and what to send to review.

Here is a call from a Worker binding:

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" },
    issued: { type: "date" },
    due_date: { type: "date", instructions: "Issue date plus the net payment days" },
    total: { type: "number", decimals: 2, unit: "USD", instructions: "Amount due including shipping" },
  },
});

The REST /ai/run endpoint works too, and you can use AI Gateway.

Pricing is $0.10 per million input tokens. There is no output token charge since the model generates no tokens.

Caveats for the preview:

  • One- and two-step calculations (totals, differences, conversions, date math) are reliable, but long dependent chains and aggregations across many rows are not yet. Extract the inputs and compute in code
  • Estimate accuracy depends on evidence in the input, such as comparables, history, or descriptive text. Without it, intervals are wide
  • Forecasts are basic. Longer horizons drift toward the average, and events not in the series, like holidays, are not anticipated
  • Intervals are trained to be calibrated, but can be too narrow on unfamiliar data
  • Answers are well formed, not guaranteed correct
  • Dates without a year are resolved from context, so set reference_date when results depend on today's date

Wrap-up

  • @cf/cloudflare/clairic is in preview on Workers AI
  • It generates no tokens and returns a typed value (or null) in one pass
  • Types: text, number, date and time types, duration, list, estimate, forecast
  • p_null and confidence help you decide what to automate
  • $0.10 per million input tokens, no output charge

If you are building pipelines that pull values out of invoices or support emails, or you are tired of LLM output drift, this one is for you!

Related posts