Cloudflare Clairic: System One model extracts typed values faster than LLMs

Cloudflare has announced a preview of Clairic, a new System One extraction model designed to pull structured data from unstructured inputs with greater speed and lower cost than traditional large language models.

Root cause

The development was inspired by TypeSafe’s earlier introduction of System One decision models, which prioritise fast, precise outputs over deliberate reasoning. Cloudflare identified that many tasks, such as calculating invoice totals or forecasting traffic, require values that cannot be pre-listed for probability scoring. This gap led to the creation of a model class that extracts, computes, estimates and forecasts direct values rather than selecting from fixed options.

Attack path

Clairic operates by processing a state of text and images through a single forward pass, avoiding the token generation phase typical of LLMs. It returns typed values, such as numbers, dates and lists, along with confidence scores and character offsets for extracted text. Because it does not generate tokens, it avoids hallucination risks associated with generative models, though it can still produce incorrect values if the input lacks sufficient evidence.

Affected versions

The model is currently available as a preview hosted on Workers AI under the identifier @cf/cloudflare/clairic. Cloudflare intends to release the model weights openly in the future. The preview supports a context window of 65,536 tokens and allows up to four images per request. No local deployment option is available until the weights are shipped.

Mitigation

• Monitor the Workers AI changelog for updates on supported field types and schema limits.
• Review confidence scores and null probabilities to determine if manual verification is needed.
• Check billing metrics, as wide schemas may incur higher input token costs than the document content.
• Validate forecast intervals, noting that multi-step arithmetic and long-range predictions may require additional scrutiny.

Why it matters

Security and operations teams should understand that not all AI tasks require generative reasoning. Extraction models offer a deterministic approach to data parsing, reducing latency and cost for structured outputs. Teams usually evaluate whether a task requires creative text generation or simple value retrieval. If the goal is to pull specific fields from documents, a System One model may provide more predictable results. Check whether your workflows can benefit from typed outputs with confidence scores rather than open-ended text, which simplifies validation and integration into downstream systems.

What to do and how to stay safe: Cloudflare

• Audit current LLM usage to identify tasks that only require extracting specific data points.
• Test the preview model with sample documents to compare latency and accuracy against existing solutions.
• Implement confidence score thresholds to automatically flag low-confidence extractions for human review.
• Monitor for the open-weights release if local deployment is a requirement for your data governance policies.

Step-by-step guide: Password Policies Examples That Reduce Password Fatigue

General security guidance from the Vector Bulletin newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

How does Clairic differ from traditional LLMs?

Clairic performs extraction in a single pass without generating tokens or reasoning, making it faster and cheaper for structured data tasks.

What types of values can Clairic return?

It returns extracted text spans, computed numbers and dates, estimated values with intervals, and forecasted future values.

Is Clairic available for local deployment?

No, it is currently hosted on Workers AI, though Cloudflare plans to release open weights in the future.