Cloudflare announced Clairic as a preview hosted on Workers AI under @cf/cloudflare/clairic. The source says a request supplies text, optionally up to four images, and named fields with types and optional instructions; the model returns typed answers in one parallel pass. These are the observed facts: the supported output types include text spans, numbers, dates, lists, estimates, and forecasts; scalar outputs carry confidence information and may be null when an answer is not found.
Thesis
Clairic makes a typed extraction interface a first-class model product, separating field extraction and bounded numeric outputs from both classification and free-form generation.
Observed facts
Cloudflare announced Clairic as a preview hosted on Workers AI under
1@cf/cloudflare/clairic.The source says a request supplies text, optionally up to four images, and named fields with types and optional instructions; the model returns typed answers in one parallel pass.
1Supported output types include text spans, numbers, dates, lists, estimates, and forecasts; scalar outputs carry confidence information and may be null when an answer is not found.
1
Mechanism
Clairic constrains the output space through declared field types and bounds rather than producing a free-form response for downstream parsing. The source describes a single forward pass without output-token generation, and positions extraction alongside decision models such as Clef and Jev and generative LLMs.
Why now
Cloudflare says recurring structured-output tasks are often sent to frontier LLMs even when their desired outputs are numbers or known field types. Clairic is presented as a specialized alternative for those tasks, with preview access already available through Workers AI.
Implications
The immediate implication is that builders can test whether a typed extractor simplifies document-processing pipelines and the validation work around them.
Typed outputs and confidence scores may support review, but do not establish correctness.
Evidence boundary
Cloudflare claims greater accuracy, speed, and lower cost than traditional LLMs for certain use cases, but the announcement gives no independent benchmarks or comparative measurements. Typed outputs and confidence scores do not establish correctness, and the company itself notes that the model can make mistakes.
Watch conditions
Watch whether independent evaluations show reliable performance across varied documents and whether the promised open-weight release arrives.