Cloudflare previews Clairic on Workers AI for typed value extraction in single pass

Cloudflare has announced a preview of Clairic, a new System One extraction model hosted on its Workers AI platform. According to first-hand research published by Cloudflare, the model is designed to extract, compute, estimate, and forecast data from unstructured inputs with higher accuracy and lower latency than traditional large language models. The company intends to release the model weights openly soon.

What problem does Clairic solve for operations?

Clairic targets tasks where developers typically use large language models to pull structured data, such as invoice totals, delivery dates, or order IDs, from unstructured text or images. According to Cloudflare, these models answer with values directly extracted, computed, or predicted, rather than generating tokens through reasoning. This approach reduces latency to under one second and eliminates output token costs, charging only for input tokens.

How does it differ from existing System One models?

Previous System One models, such as TypeSafe’s Jev and Cloudflare’s Clef, are decision models that score a fixed list of options against probabilities. Clairic is the first System One extraction model, handling answers that cannot be listed in advance. It returns typed values like names, dates, and lists with confidence scores. According to MarkTechPost, it processes requests in one forward pass without reasoning, contrasting with the multi-second processing time of reasoning-based large language models.

What are the technical limits and costs?

The model supports a context window of 65,536 tokens and up to four images per request. According to MarkTechPost, the cost is $0.10 per million input tokens with no charge for output. Fields that exceed the token limit for a single pass are split into parallel passes, each seeing the full input. This means wide schemas can incur higher costs than the source document itself, as schema tokens are billed alongside input tokens.

What are the performance risks?

Cloudflare’s tests showed sub-second extraction times, but the model has limitations. According to MarkTechPost, multi-step arithmetic is unreliable, and some forecast intervals may miss actual values. In a specific test, four of twenty-one next-day forecasts fell outside the model’s 80% confidence interval. The model returns null when it fails to find an answer, helping teams identify missing data rather than hallucinated content.

Questions to ask your suppliers

Security and engineering teams should evaluate how this model handles sensitive data within their specific workflows. Check whether the supplier validates input types strictly to prevent injection attacks or schema confusion. Ask if the model’s confidence scores are calibrated for your industry’s risk tolerance. Determine who inside the organization needs to review the model’s null outputs and error rates before production deployment. Ensure that logging mechanisms capture both input and output for audit trails, especially when financial or personal data is processed.

Background: Large language models

Large language models vs machine learning differs in scope and output. Machine learning is the broad discipline of algorithms learning from data. Large language models are a specific type of deep learning model focused on understanding and generating human language by predicting the next word in a sequence.

Read the full guide: Large Language Models vs Machine Learning: Core Differences Explained

What to do and how to stay safe: Cloudflare

• Review current large language model usage for simple extraction tasks to identify candidates for replacement with specialized models.
• Validate the accuracy of extracted values against ground truth data before integrating the model into critical business workflows.
• Monitor costs closely, as wide schemas may increase input token usage through parallel processing passes.
• Establish clear thresholds for confidence scores to determine when human review is required for ambiguous outputs.

Step-by-step guide: Browser Security Settings for Small Business Teams

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

Frequently asked questions

Is Clairic available for local deployment?

No, it currently runs on Cloudflare-hosted GPUs via Workers AI, though the company plans to release open weights soon.

Does Clairic charge for output tokens?

No, according to MarkTechPost, it charges only for input tokens at a rate of $0.10 per million tokens.

Can Clairic handle images?

Yes, it supports up to four images per request alongside text inputs within its context window.