Cloudflare previews Clairic extraction model that returns structured values in a single forward pass

Clairic is a new category Cloudflare calls a System One extraction model: a request contains only the input (text plus up to four images) and a set of typed fields, and the model returns results in parallel in a single forward pass without generating tokens, so billing is input-only and responses typically arrive within a second. Supported types include text (with character offsets and confidence), number (respecting integer or decimal units), date, time, datetime, duration, list, plus estimate (returning a median and 80% interval over a [min,max] range and step grid) and forecast (step-by-step series prediction). Each scalar answer carries p_null and confidence, and the server splits large field sets into parallel branches then merges them. For teams currently using frontier LLMs for structured output, the shift is that output is no longer a string: invoice numbers, dates, and amounts come back as verifiable typed values, missing fields return null instead of being fabricated, and amounts or day counts can be computed from prices, tax, or two dates in the input. In examples, 30 fields cost about the same as 3. The most direct impact is on developers running order parsing, support tickets, or receipt entry on Workers: they can call it via env.AI.run and reuse their existing Cloudflare account, though the model is currently preview-only, with weights planned for release soon.