N環球 
Cloudflare unveils 'Clairic' extraction model: no text generation, values in a single pass
Cloudflare announced on October 19 that its Workers AI platform is previewing Clairic, a System One extraction model that returns typed values instead of generated text and charges input tokens only.
Why it matters
Large language models are slow and expensive for structured extraction tasks, and they can hallucinate numbers. Clairic takes a different route: a single forward pass returns the answer directly, with no generated text — faster, cheaper and easier to audit, exactly what enterprise document pipelines (invoices, claims, audits) have been missing.
Full report
Cloudflare announced on October 19 on its official blog that its Workers AI inference platform has begun previewing Clairic (model ID: @cf/cloudflare/clairic). The company calls it the first member of a new category, "System One extraction models" — System One referencing psychologist Daniel Kahneman's "fast thinking": answering in one pass without reasoning or generating tokens.
Clairic targets a specific developer workload: finding an invoice total, computing how late a delivery was, estimating a repair bill. It supports four outputs: extract (exact spans from the input), compute (e.g., a Net-30 due date), estimate (median with an 80% interval), and forecast (up to 64 future steps). Answers come back typed, with character offsets and confidence scores, or null when the input doesn't determine them.
Pricing is the selling point: input tokens only, $0.10 per million tokens, with a 65,536-token context window. Cloudflare says an internal test showed a "Clairic + decision model Clef + LLM" claims pipeline ran 69% faster and cost 49% less than an all-LLM version; on public TSA daily-screening data, forecasts averaged 3.3% error. The company also admits multi-step arithmetic and long-range forecasting remain weak, and says open weights are coming "soon".
This is Cloudflare's second System One expansion this month — decision model Clef launched on October 1. Together with TypeSafe's Jev in September, models that don't generate text are becoming AI infrastructure's new battleground.
💬 Comments
No comments yet. Be the first!