Cloudflare previews Clairic, a typed extraction model on Workers AI
Cloudflare released a preview of Clairic, a non-generative model that returns typed values with confidence scores instead of free-text LLM output.
Cloudflare has put Clairic in preview on Workers AI as @cf/cloudflare/clairic. It is an extraction model that returns typed fields instead of generating text. If you pay an LLM to pull invoice totals, dates or order IDs out of documents and then validate its JSON, this targets that job. Cloudflare's October 19 post says open weights are coming "soon", with no date given.
You send a state (text plus up to four images) and a map of named, typed fields: text, number, date, duration, list, estimate and forecast. Every field is answered in one forward pass. Each scalar answer comes back with a confidence score and p_null, the probability that the value isn't in the input. When the input doesn't answer a field, the value is null. Extracted text includes character offsets, so you can trace each value to the source. According to Cloudflare's changelog, it costs $0.10 per million input tokens, with no output charge.
The limits are stated up front. The changelog says one- and two-step arithmetic is reliable, but long chains and sums across many table rows are not. Its advice is to do that math in code. Forecasting is described as basic, and the changelog warns that every answer is "well formed, not necessarily correct."
The speed and cost figures come from Cloudflare's own test. Help Net Security reports that a claims pipeline built around Clairic finished in 19.3 seconds against 61.7 seconds for an all-LLM version, at 49% lower cost. No independent benchmark has been published that we found. For bookkeeping or support-triage pipelines, it's worth testing on your own documents. Route low-confidence fields to review.