What changed
On October 19, 2026, Cloudflare announced Clairic, the first hosted preview of its System One extraction-model class, available through Workers AI as @cf/cloudflare/clairic. Instead of generating prose tokens, it answers named fields with typed text spans, numbers, dates, durations, lists, estimates or forecasts. Outputs include confidence, p_null and probability intervals for some numeric tasks. Cloudflare intends to release open weights later, but the preview is currently hosted and the weights are not yet available.
Why it matters
A common application pattern asks a general LLM to extract invoice fields, parse dates, classify a claim, then format JSON. Clairic attempts to make the extraction portion a cheap typed operation with spans tied to input and predictable output shapes. That could simplify ingestion pipelines and reduce latency, but type correctness does not mean factual correctness; operators still need validation, confidence thresholds and deterministic arithmetic for important decisions.
A schema-bound extraction API
A request sends state text and up to four optional images, plus a map of fields. Types include text spans with character offsets, number, date, time, datetime, duration, list, bounded estimate and multi-step forecast. The API returns declared types or null rather than free-form generated text. It can split large field maps into parallel passes that share the input.
One-pass economics and reported tests
Cloudflare says Clairic is billed for input processing without generated output tokens, and most calls take less than a second. In its own 21-day TSA passenger-count experiment, mean absolute percentage error was about 3.3%, with 17/21 observations inside its nominal 80% interval. In a staged customer-support pipeline combining Clairic, Clef and GLM-5.3, Cloudflare reports 19.3 seconds and $6,032 per million claims versus 61.7 seconds and $11,740 for an all-LLM pipeline. These are vendor-run demonstrations, not independent benchmarks or guaranteed savings.
Where not to trust the preview
Cloudflare says one- and two-step calculations work better than long arithmetic chains; estimates require contextual evidence, interval calibration can drift on unfamiliar data, and forecasting is not a replacement for a dedicated time-series model. A syntactically valid answer may still be wrong. Use p_null, confidence and intervals to route uncertain outputs to review, and use application code for deterministic financial arithmetic.
Deployment and open-weight boundaries
The currently documented route is a hosted Workers AI preview at @cf/cloudflare/clairic via Worker env.AI.run or REST. Open-weight release is an intention, not a shipped artefact. Test latency, cost and accuracy against your own documents before switching from general LLM extraction.