Cloudflare previews Clairic, a System One extraction model for structured data

• Cloudflare releases the preview of Clairic, a System One extraction model hosted on Workers AI at @cf/cloudflare/clairic, with open-weights release planned.
• In a cascade with Clef and GLM-5.3, Clairic reduces total latency from 61.7s to 19.3s (69% faster) and cost from $11,740 to $6,032 per 1M requests (49% cheaper) compared to an all-LLM pipeline.
• The model performs single-pass extraction, computation, estimation, and forecasting without token generation, returning typed values with confidence scores and p_null to prevent hallucinations.
• TSA airport security data forecasting showed a 3.3% average error over 21 days, with 17 of 21 days falling within the 80% confidence interval.

Capabilities and Limitations

• Supports text, number, date, time, datetime, duration, list, estimate, and forecast types from text and up to 4 images.
• Reliable for 1-2 step arithmetic (sums, differences, date conversions) but struggles with complex dependent steps or table aggregation.
• Estimation requires evidence in the input; forecasting is limited to general pattern matching and cannot replace dedicated time-series models for unseen events like holidays.