Close-up photo of a green circuit board with traces and small components

· Translated from the Spanish original

AI and data

Cloudflare launches Clairic and charges $0.10 per million tokens without generating text

The model, in preview on Workers AI, reads the input once and returns one value per field, or null when the input does not settle it. No output charge.

What happened

  • On October 19, 2026, Cloudflare published the changelog entry for Clairic, a model that extracts typed values, available on Workers AI as a preview.
  • The caller sends a state and a schema of typed fields. The model returns one value per field: a number, a date, a span of text or a probability distribution.
  • According to Cloudflare’s document, every answer is a valid value of the requested type, or null when the input does not determine it.
  • The price is $0.10 per million input tokens. Because the model generates no text, there is no charge for output tokens.

Why it matters

  • Clairic does not write or reason out loud: it reads and decides. For a small business pulling amounts and dates out of invoices, emails or records, spending depends only on how much the model reads, not on how much it writes.
  • The null is the useful part. A system that can say “I don’t know” avoids filling a field with an invented figure, a mistake that in accounting or a clinical record is found only after it has cost money.
  • The source is the company’s own changelog. There is no independent test yet, let alone one run on Chilean Spanish documents, with their tax ID formats, dates and peso amounts.

The number

$0.10 per million input tokens, and zero for output tokens.

Context

Cloudflare had already introduced Clef, models that choose instead of write, and Microsoft launched Decision-1, which charges only for input. Clairic follows the same logic of models that produce a structured decision rather than a paragraph, but it aims at extracting and computing values rather than choosing between options.

What’s next

  • It is a preview: the document consulted sets no date for general availability.
  • No announced timeline for specific Spanish support or for publishing the weights.
  • Before adopting it, measure how often it answers null on your own data: that is where the model’s real quality shows.

Bottom line

A model that can answer “null” is the opposite of a chatbot that always replies. That ability to abstain is exactly what the study of the abstention and hallucination circuit showed general models struggle with.

Sources

Edited by Rodrigo Cornejo. How we select and verify: who writes these notes.

← All notes