Cloudflare launches Clairic, a System One extraction model that returns typed values, dates and forecasts with confidence in one forward pass, in preview on Workers AI at $0.10 per million input tokens

Background and pain point

Many structured-output calls to frontier LLMs ask the same questions: what is the invoice total, when was it delivered, how late is it, what will the repair cost. LLMs can answer, but they emit invalid JSON, misread digits while generating, and return excerpts that cannot be traced back to the source, so teams stack parsing, validation, retries and spot checks, and still pay for reasoning and output tokens on every field. Borrowing TypeSafe's "System One" idea (fast, precise, typed), Cloudflare adds the first extraction model next to decision models like Jev and Clef.

Architecture and mechanism

Clairic generates no tokens: one parallel forward pass answers every field, so only input is billed. A request is a state (text plus up to 4 images) and a map of typed fields: text, number, date/time/datetime, duration, list, estimate and forecast. Text values come back as character-offset spans with confidence; every scalar carries p_null and returns null instead of a guess when the value is missing; estimates return a median and an 80% interval (p10 to p90), and forecasts give one interval per step. Field count is unlimited; the server splits large requests into parallel passes and merges them.

Benchmarks

All figures are Cloudflare's own; no third-party evaluation is public yet. In a damage-claim pipeline, the Clairic to Clef to GLM-5.3 cascade took 19.3 s end to end versus 61.7 s with GLM-5.3 for every step, about 3.2x faster; extraction alone took 0.46 s versus 16.1 s; cost per million claims was $6,032 versus $11,740, about 49% less. A 21-day forecast of TSA checkpoint counts averaged 3.3% error, with 17 of 21 days inside the 80% interval. Cloudflare says long chains of arithmetic and aggregation across many rows are not reliable yet, and intervals can run narrow on unfamiliar data.

Takeaways for builders

Call env.AI.run("@cf/cloudflare/clairic", { state, fields }) from a Worker or use the REST API. The docs list a 65,536-token context, vision support and $0.10 per million input tokens. It fits invoice, receipt, ticket and email extraction plus estimates with uncertainty; Cloudflare suggests letting Clairic read values, letting code do heavy math, and handing typed results to Clef or an LLM. It is an early preview checkpoint, and Cloudflare says open weights will follow.