Cloudflare Clairic returns typed values, not tokens

What an extraction model is

Decision models such as Cloudflare's Clef pick among options you list up front. Many practical questions have no enumerable option set: an invoice total, a delivery date, days late, an expected repair cost. Clairic handles those. You send a state and a schema of typed fields, and one parallel pass returns a value per field (Cloudflare Changelog, Cloudflare Blog).

Which types it returns

• text: an exact span of the input with character offsets
• number, date, time, datetime, duration: under the rules of each type
• list: matching items in input order
• estimate: a distribution inside your min-max bounds, with a median and an 80% interval
• forecast: one estimate per future step, up to 64 steps

Every scalar answer includes p_null (probability that the input does not determine the value) and confidence. You can pass up to four images. Cloudflare says open weights are planned.

How much faster than an all-LLM path

The blog measures a customer-support cascade: Clairic + Clef + a generative model versus an all-LLM pipeline. The cascade finished in 19.3 seconds versus 61.7, about 3.2x faster and 49% cheaper. Clairic and Clef together finished extraction and the decision in under a second.

Pricing is $0.10 per million input tokens. Because Clairic generates no tokens, there is no output charge.

Where it still falls short

One- and two-step arithmetic is reliable. Long dependent chains and aggregations across many table rows are not. Forecasting is basic: farther steps flatten toward the average, and holidays are not anticipated. Intervals are calibrated on most held-out sets but can run narrow on unfamiliar data. A well-typed answer is not automatically a correct one.