Cloudflare Releases Clairic, a System One Extraction Model on Workers AI

Cloudflare has put a preview of Clairic on Workers AI, and the pitch is simple: stop asking big language models to do every little data-cleaning job. Clairic reads text or images and returns typed values, not generated prose. Think totals, dates, IDs, estimates, forecasts. It answers in one forward pass, with no reasoning chain and no output tokens to pay for.

That makes it part of a three-layer idea Cloudflare is pushing. Extraction models turn messy input into typed values. Decision models turn those values into choices. LLMs handle the part humans actually read: the words. Clairic sits in the first layer, alongside Cloudflare’s own Clef and TypeSafe’s Jev, but it is aimed at answers you cannot pre-list in advance.

The model accepts text plus up to four images, with a 65,536-token context window. Supported field types include text, number, date, time, datetime, duration, list, estimate and forecast. Every scalar answer comes with p_null and a confidence score. Text spans include character offsets. Estimates and forecasts include a median and an 80% interval. There is no field limit, though wide schemas can force parallel passes and push token use well past the source document itself.

Cloudflare says Clairic handled a support email by adding price and tax to $701.10 and calculating 11 days since delivery, even though neither figure was written there. On a skewed receipt photo it read all seven present values and returned null for the missing discount. In forecasting tests, it modeled 21 days of TSA checkpoint counts with 3.3% average error, and 17 of 21 actual counts landed inside the 80% interval. The misses clustered around Columbus Day weekend.

The clearest proof point is a damaged-product claims pipeline. Cloudflare split the work across Clairic, Clef and GLM-5.3, then compared that with GLM-5.3 doing all three steps. The split version finished in 19.3 seconds versus 61.7 seconds for the all-LLM setup. Extraction alone dropped to 0.46 seconds from 16.1 seconds, and the decision step fell to 0.22 seconds from 21.9 seconds. Cost per million claims went from $11,740 to $6,032. The reply stage still ate most of the time, at 18.6 seconds out of 19.3.

Clairic is live as a preview at @cf/cloudflare/clairic through env.AI.run() or the REST API, and AI Gateway can hit the same endpoints. It costs $0.10 per million input tokens, with no charge for output tokens. Weights are planned, but not out yet.