Cloudflare has launched Clairic, a new model that extracts structured data from unstructured text and images in a single forward pass, reducing costs and latency compared to traditional LLMs. The model is now available in preview on Workers AI.
TL;DR
- Cloudflare's Clairic model extracts typed values from raw inputs without generating tokens or reasoning through steps.
- The model is part of Cloudflare's System One concept, which contrasts fast, single-pass models against slower, deliberative System Two LLMs.
- Clairic is positioned as a cost-effective and efficient alternative for teams spending LLM tokens on extraction workloads.
What happened
Cloudflare has released a preview of Clairic, a new model designed to extract structured data from unstructured text and images. The model is available on Workers AI as @cf/cloudflare/clairic.
Clairic pulls numbers, dates, spans of text, and probability distributions from raw inputs without generating tokens or reasoning through steps. Open weights are planned for a later release.
The model is part of Cloudflare's System One concept, which contrasts fast, single-pass models against slower, deliberative System Two LLMs. Clairic handles extraction, while TypeSafe's Jev handles classification decisions.
Why it matters
Clairic offers a concrete alternative for teams spending LLM tokens on invoice parsing, receipt OCR, and similar extraction workloads. The model reduces costs and latency compared to traditional LLMs.
Cloudflare benchmarked a three-step claim-resolution pipeline using Clairic, Clef, and GLM-5.3 against an all-GLM-5.3 baseline. The cascade completed in 19.3 seconds versus 61.7 seconds, a 69% latency reduction. Estimated cost per million claims fell from $11,740 to $6,032.
The release signals a shift toward specialized, non-generative models for structured-data tasks that currently consume disproportionate LLM compute. However, the long-term durability of the System One framing remains to be seen.
Key facts
- Clairic extracts typed values, including text spans with character offsets, computed numbers and dates, estimates with 80% intervals, and forecasts with per-step confidence bands.
- The model accepts images and returns order numbers, timestamps, line items, subtotals, taxes, tips, and null discount fields with calibrated confidence scores.
- For estimation tasks, Clairic returns a distribution over a declared range with a median, 80% interval, and optional percentiles. A real-estate example produced a sale-price estimate of $418,000 with a $370,000–$470,000 interval.
- For forecasting, the model ingested 21 days of TSA checkpoint traveler counts and predicted the next day within 3.3% average error; 17 of 21 actuals fell inside the 80% interval.
- Cloudflare benchmarked a three-step claim-resolution pipeline using Clairic, Clef, and GLM-5.3 against an all-GLM-5.3 baseline. The cascade completed in 19.3 seconds versus 61.7 seconds, a 69% latency reduction. Estimated cost per million claims fell from $11,740 to $6,032.
- The preview carries caveats, including multi-step arithmetic degradation, the need for evidence in the input for estimates, and calibration that can run narrow on unfamiliar distributions.
- Developers can call Clairic from Workers via env.AI.run() or from any environment through the REST API. Documentation and an API explainer are published.
Context
Clairic is part of Cloudflare's broader strategy to offer specialized, non-generative models for structured-data tasks. The model is positioned as a cost-effective and efficient alternative for teams spending LLM tokens on extraction workloads.
The release signals a shift toward specialized models for structured-data tasks that currently consume disproportionate LLM compute. However, the long-term durability of the System One framing remains to be seen.
Clairic is available in preview on Workers AI, with open weights planned for a later release. The model name combines 'Clair' (clear, clerical, clairvoyance) with Cloudflare's 'AI Regression' initials.