Cloudflare previews Clairic, an extraction model that returns typed values with intervals in one forward pass
Synopsis
Cloudflare has previewed Clairic on Workers AI, extending so-called System One models from decisions to extraction: given text and a map of typed fields, it returns extracted, computed, estimated or forecast values in a single parallel forward pass, each with confidence and an 80% interval. Structured-output calls previously went to frontier LLMs that generate tokens, requiring parsing, validation and retries. In one support-claim pipeline, Clairic plus the decision model Clef plus an LLM finished in 19.3 seconds, against 61.7 seconds when GLM-5.3 did every step.
Interpretation
Structured field extraction now has a dedicated model class instead of relying on a general LLM to generate text that is then parsed. Such requests previously went repeatedly to frontier LLMs, whose string outputs needed layers of parsing, schema validation, retries and spot checks, with reasoning and output tokens paid on every field. On the same support email, policy and prompts, a pipeline of Clairic extraction, Clef decision and GLM-5.3 reply took 19.3 seconds end to end, versus 61.7 seconds when GLM-5.3 performed all three steps, about 3.2 times faster.
A request is a piece of text, optionally with up to four images, plus a map of fields; each field declares a type and optional plain-language instructions, and one parallel pass returns one typed answer per field. The number of fields is unlimited: when they do not fit beside the input in one pass, the server splits them into parallel passes that each see the whole input and then merges the answers. On the example support email, the model summed 649.00 and 52.10 dollars into 701.1 although the total is not written, derived 11 days from two dates, and returned null for an absent tracking number instead of a plausible guess.
Numeric answers come with a distribution, a median and an 80% interval, and the interval width reflects how much the input informs the prediction. Asking an LLM for an estimate usually yields a single number, so downstream code cannot tell a confident 180 from an unconfident one, while decision models can only approximate estimates with coarse buckets. Forecasting next-day TSA checkpoint counts for 21 consecutive days, Clairic was off by 3.3% on average and 17 of the 21 actual counts fell inside its 80% interval, with three of the four misses on the Columbus Day weekend.
Perspective
The demonstrated setting is a hosted preview endpoint on Workers AI, taking text plus up to four optional images and returning outputs confined to the declared types and given bounds; model weights are not yet open, and the release says they are planned. The release recommends Clairic for non-critical general pattern matching rather than as a replacement for time-series models trained on specific data.
The release calls the endpoint an early checkpoint: multi-step arithmetic and aggregation across many table rows remain unreliable, and it suggests letting the model predict values while code does the math; intervals are trained to be calibrated but can run narrow on unfamiliar data; forecasting suits short-term general patterns and cannot account for events the series does not show, such as holidays.