Introducing Clairic, a new System One class of extraction models

Cloudflare has introduced Clairic, the first extraction model in a new System One class, available now in preview on Workers AI as @cf/cloudflare/clairic. Unlike traditional LLMs, Clairic processes inputs in a single forward pass to extract, compute, or estimate values such as numbers, dates, or lists directly from unstructured text or images. The model returns typed outputs with confidence scores and avoids token generation, reducing latency and cost while eliminating common LLM failure modes like hallucinations or invalid formats.

The announcement follows TypeSafe’s recent introduction of System One decision models like Jev, which prioritize speed and precision over deliberate reasoning. Clairic expands this concept by focusing on structured-output extraction tasks that LLMs handle inefficiently, such as retrieving invoice totals, delivery dates, or repair cost estimates. By avoiding multi-step reasoning, Clairic achieves sub-second response times and charges only for input tokens, making it suitable for high-volume workflows where accuracy and speed are critical.

Clairic’s design includes support for scalar values with confidence intervals, null probability assessments, and optional percentile breakdowns for estimates. It accepts up to four images alongside text and processes fields in parallel passes if needed, merging results without additional reasoning steps. Outputs are strictly typed, with each field’s answer accompanied by a confidence score and a probability that the value is absent from the input, enabling downstream systems to evaluate reliability before acting.

In practical terms, Clairic replaces complex extraction pipelines that rely on LLMs, JSON parsing, and manual validation. For example, given a support email with a damaged product and payment details, Clairic extracts the order number, arrival date, and tax amount while flagging missing fields like discounts. Similarly, it can process receipt images with imperfect text to retrieve line items, totals, and tax amounts, demonstrating its ability to handle real-world data with high accuracy and minimal overhead.