Cloudflare Preview Clairic Model Extracts Data Faster and Cheaper Than LLMs
Cloudflare’s new extraction model processes documents in seconds at half the cost of large language models, eliminating token generation overhead.
Key points
- Cloudflare released a preview of Clairic, an extraction model that pulls typed values from text and images.
- In tests, Clairic pipelines finished in 19.3 seconds and cost 49% less than LLM-based alternatives.
- The model charges only for input tokens and returns confidence scores for each extracted field.
Cloudflare released a preview of Clairic, a specialized extraction model designed to pull structured data from messy text and images. According to Help Net Security, the model operates without generating output tokens, addressing a common inefficiency in current large language model workflows. This approach targets routine developer tasks such as reading invoices to extract order numbers, totals, and due dates.
Traditional large language models often fail at these specific tasks by emitting broken JSON, misreading digits, or returning excerpts without source pointers. Developers typically add parsers, schema checks, and retries to mitigate these errors, paying for output tokens on every field. Clairic runs on Workers AI and allows developers to specify named fields with types like number, date, or list.
Technical Performance and Cost
In a test involving a damaged-product support claim, Cloudflare compared Clairic against a pipeline using GLM-5.3 at every step. The Clairic-based pipeline finished in 19.3 seconds, while the LLM version took 61.7 seconds. The Clairic workflow also cost 49% less, totaling $6,032 per million claims compared to $11,740 for the LLM cascade.
Clairic extracted the required facts in 0.46 seconds during the claims test. Cloudflare’s Clef model then selected a refund or repair option in 0.22 seconds. The GLM-5.3 model required 38 seconds for these two steps alone. Both pipelines used GLM-5.3 to write the final reply, which consumed 18.6 seconds of the faster cascade’s total time.
Accuracy and Reliability
Both pipelines correctly identified facts and decisions in the runs Cloudflare reported. However, Clairic provided a steady repair estimate, whereas GLM-5.3 varied its output between $250 and $220 across runs. Other models, including Gemma 4 and Qwen, failed to derive the repair cost when run as all-LLM pipelines.
When asked for data not present in the input, such as a missing tracking number, Clairic returned null. It provided a fallback guess with a 0.42 confidence score when appropriate. Scalar answers include confidence scores and p_null values, indicating the probability that a value is missing from the input source.
Forecasting Capabilities
Cloudflare tested Clairic’s ability to predict future values by asking for the next day’s TSA checkpoint count at U.S. airports over 21 days. The forecasts missed by 3.3% on average. Seventeen of the 21 actual counts landed inside Clairic’s 80% confidence interval, demonstrating the model’s utility for estimation tasks beyond simple extraction.
Who is affected
Developers who routinely use large language models for data extraction from invoices, emails, or forms are the primary audience. Teams currently stacking parsers and retries to handle LLM hallucinations or format errors may benefit from switching to a dedicated extraction model. Cloudflare plans to publish the model weights, allowing for broader integration and potential local deployment.
What should you do?
Security and technology professionals should evaluate whether their current LLM usage involves high-volume, structured data extraction tasks. If teams are paying for output tokens on simple fields, switching to an extraction model could reduce costs and latency. Developers should test Clairic’s ability to handle their specific document formats and confidence thresholds.
Background: Large language models
Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.
What to do and how to stay safe: Cloudflare
- Review current LLM workflows to identify tasks that involve extracting specific fields from documents or images.
- Test extraction models like Clairic against existing LLM pipelines to compare latency, cost, and accuracy.
- Implement confidence score thresholds to automatically flag low-confidence extractions for human review.
- Monitor vendor updates for model weight releases and integration documentation for Workers AI.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What is Clairic?
Clairic is an extraction model released by Cloudflare that pulls typed values from text and images without generating output tokens.
How much does Clairic cost compared to LLMs?
In Cloudflare’s tests, Clairic cost 49% less than an LLM-based pipeline, at $6,032 versus $11,740 per million claims.
How accurate is Clairic at forecasting?
In a 21-day test predicting TSA checkpoint counts, Clairic’s forecasts missed by 3.3% on average, with 17 actual counts inside its 80% interval.
