Cloudflare releases Clairic, a no-reasoning extraction model, on Workers AI

Curated by agents · checked by the desk

Cloudflare released Clairic in preview on Workers AI on October 19, 2026, as @cf/cloudflare/clairic. The company calls it the first of a new “System One” class of extraction models: it reads an input and returns typed values in a single forward pass, with “no reasoning and no output tokens,” and it charges only for input. In a damage-claim pipeline Cloudflare built, moving extraction to Clairic cut cost to $6,032 per million claims from $11,740 and end-to-end latency to 19.3 seconds from 61.7 seconds.

Developers call it from a Worker with env.AI.run() or from anywhere through the REST API. Requests declare fields by type: text (returned as a span with character offsets), number, date, time, datetime, duration, list, estimate and forecast. Estimates come back as a median with an 80% interval (p10 to p90) inside declared bounds, and forecasts return an interval per step. Outputs carry confidence scores, and the input can include up to four images. Cloudflare says the model typically responds in under a second. In the claims pipeline, Clairic extracted the customer name, order ID, amount, days since delivery, damage list and a repair-cost estimate in 0.46 s. Clef, a System One decision model, chose refund or repair in 0.22 s, and GLM-5.3 (@cf/zai-org/glm-5.3) wrote the reply in 18.6 s. The baseline ran GLM-5.3 for all three steps, spending 16.1 s to extract and 21.9 s to decide. Cloudflare puts the result at 69% less time and 49% lower cost. On a 21-day forecast of TSA daily passenger counts, the company reports an average error of 3.3%, with 17 of 21 actual counts inside the 80% interval.

The “System One” label comes from Kahneman’s Thinking, Fast and Slow, with LLMs such as GPT, Claude, Gemini and DeepSeek cast as System Two. Cloudflare pitches Clairic against the parsing, schema validation, retries and spot checks that teams layer on top of LLM extraction, and notes that Clairic’s repair estimate stayed identical across runs while GLM-5.3 returned $250, $250 and $220. The company lists preview limits: one- or two-step arithmetic is reliable, but longer calculations and aggregation across table rows are not yet, and the authors suggest letting code do the math. Estimates work best with supporting evidence in the input, calibration is “measured, not guaranteed,” intervals may be too narrow on unfamiliar data, and forecasting does not replace dedicated time-series models.

Several details are absent. The post gives no per-token price for Clairic, only pipeline totals. It states no parameter count, architecture, maximum input length or supported languages. Cloudflare describes the endpoint as “an early checkpoint” and names no general-availability date or rate limits. On weights, it says only “We intend to release the model open-weights soon,” with no date or license. The claim that Clairic is more accurate than LLMs for certain uses comes without a dataset or accuracy figures.

Source: Cloudflare, October 19, 2026.

Why it matters under load: Input-only billing removes output-token spend from extraction steps, but without a published per-token rate or rate limits, the 49% saving cannot yet be modeled at production volume.