Cloudflare Preview Model Cuts Extraction Costs 49 Percent Without Output Tokens

Cloudflare released a preview of Clairic, an extraction model designed to pull structured data from unstructured sources like invoices and support tickets. This development impacts organizations that rely on large language models for routine data processing, as the new tool offers a faster and cheaper alternative for specific extraction tasks. The model runs on Workers AI and charges only for input tokens.

By the numbers

• 19.3 seconds: The time required for a Clairic-based pipeline to process a damaged-product support claim in Cloudflare’s tests.
• 49 percent: The reduction in cost for the extraction pipeline compared to a version using a large language model at every step.
• 0.46 seconds: The time Clairic took to extract facts from a claim, while a comparison model took 38 seconds for the same steps.

What the figures mean

According to Help Net Security, Cloudflare’s tests demonstrated that Clairic can handle messy text and images without the overhead of token generation. In a sample test, the model correctly calculated totals and identified missing values, returning null for absent data. It also provided confidence scores for its estimates. The model outperformed large language models in consistency, providing steady repair cost estimates where other models varied.

What should you do?

Managed service providers should evaluate their current workflows for repetitive data extraction tasks, such as invoice processing or form filling. Identify processes where large language models are used solely to extract specific fields, as these are prime candidates for optimization. Review vendor pricing models to understand the cost differences between input-based and output-based charging structures. Assess the accuracy requirements of your clients’ data pipelines to determine if a specialized extraction model meets their needs.

Operational impact

Teams usually need to audit existing data pipelines to identify where large language models are performing simple extraction tasks. Check whether current vendors support specialized models for structured data, as this can reduce latency and costs. Inform your engineering teams about new extraction tools that do not rely on output tokens, which may change how you budget for API usage. Discuss with suppliers whether their platforms can integrate these new models to improve processing speeds and reliability for client data.

Background: Large language models

The difference between large language models and small language models lies in their parameter counts, training data volumes, and memory footprints. Large models handle complex, open-ended reasoning across many domains, while small models execute narrow tasks efficiently on local hardware with predictable latency and lower resource overhead.

Read the full guide: Difference between large language models and small language models for engineering choice

What to do and how to stay safe: Cloudflare

• Review current large language model usage to identify tasks that involve only extracting specific data fields from documents or images.
• Test the new extraction model on non-critical workflows to compare processing speed and accuracy against your current setup.
• Monitor vendor announcements for updates on model weights and integration capabilities, as these affect long-term deployment plans.
• Consult with development teams to understand how input-only pricing models might impact overall operational costs for data processing tasks.

Step-by-step guide: Shadow AI Explained: 8 Questions Answered for Secure Operations

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

Frequently asked questions

How does Clairic differ from traditional large language models?

Clairic extracts typed values from text and images in one pass without generating output tokens, while large language models generate text for every field.

What was the cost difference in Cloudflare’s tests?

The pipeline using Clairic cost 49 percent less than the version using a large language model at every step, according to Cloudflare’s reported figures.

Does Clairic handle missing data?

Yes, the model returns null for values not present in the input and provides confidence scores for its estimates and extracted text.