Cloudflare released Clairic on October 19, 2026 — a preview model on Workers AI that does not write text at all. It reads unstructured text and images and returns typed values: extracted, computed, estimated or forecast, each with a confidence score. Why it matters: a large share of expensive frontier-LLM calls are really just “read this invoice and give me the total and due date as JSON.” Clairic is built to do only that job, in one pass, for $0.10 per million input tokens — with no output tokens to pay for.
October 20, 2026 — Tested Intelligence News Desk
What happened
Cloudflare announced Clairic in a blog post on October 19, calling it the first “System One extraction model.” It follows Clef, Cloudflare’s System One decision model, which picks among fixed options and returns probabilities. Clairic covers the other common case: answers that cannot be listed in advance, like an amount, a date or a delivery delay. The model is live in preview as @cf/cloudflare/clairic on Workers AI, callable from a Worker or the REST API, and Cloudflare says it intends to release the weights openly after further checkpoint work, without giving a date.
How it works
You send a state — text plus up to four images — and a map of named, typed fields. In a single forward pass, with no reasoning and no generated tokens, Clairic returns one typed answer per field: a text span with character offsets, a number, a date, a duration, a list, an estimate with a median and an 80% interval, or a forecast with an interval per step. Every scalar answer carries a confidence score and a p_null — the model’s probability that the value simply is not in the input — so a missing tracking number comes back as null instead of a plausible-looking guess. The model documentation lists a 65,536-token context window and pricing of $0.10 per million input tokens, with no output charge.
In Cloudflare’s own damaged-product claim test, a pipeline using Clairic for extraction, Clef for the decision and GLM-5.3 only for the written reply finished in 19.3 seconds and cost 49% less than an all-LLM version, which took 61.7 seconds. On 21 days of TSA airport screening data, Clairic’s next-day forecasts were off by 3.3% on average, with 17 of 21 actual counts landing inside its 80% interval.
The honest caveats
Cloudflare is unusually upfront about the limits: multi-step arithmetic and aggregating across many table rows are weak in a one-pass model, estimates need evidence in the input, intervals can run narrow on unfamiliar data, and forecasting is basic pattern-matching — it missed the Columbus Day travel surge because holidays are not in the series. A well-formed answer, the company notes, is not necessarily a correct one; the confidence scores exist so your code can decide what to automate and what to send for review.
Why it matters
Extraction is where agent pipelines quietly burn money: parsers, schema checks and retries stacked on top of LLM calls that were never designed to return a number. If a $0.10-per-million-input model can take that layer over reliably, the expensive model only gets called for the step that actually needs reasoning or writing. That is the “everything AI, tested” angle worth watching — not a bigger model, but a narrower one that makes the rest of the stack cheaper. Developers can try the preview on Workers AI today.
Frequently asked questions
Is Clairic an LLM?
No. Cloudflare classifies it as a System One extraction model: it generates no tokens and returns typed values in a single pass, rather than writing text.
How much does Clairic cost?
$0.10 per million input tokens on Workers AI. There is no output-token charge, because the model produces no output tokens.
Are the weights open?
Not yet. Cloudflare says it intends to release Clairic open-weights after further checkpoint work, but has not given a date. For now it runs only on Cloudflare’s endpoint, in preview.
Sources: Cloudflare Blog, Unite.AI, Help Net Security