Editorial illustration for Cloudflare Adds Clairic System One Model to Workers AI for Data Extraction
Cloudflare Adds Clairic to Workers AI for Data Extraction
Cloudflare pushed a preview of Clairic onto Workers AI this week, pitching it as a fix for a job developers keep handing to the wrong tool. Pulling a total off an invoice or a date out of a scanned form doesn't need a chatbot that writes sentences. It needs a model that reads the input and spits out a typed value: a number, a date, an ID. Clairic does that in one forward pass, no reasoning steps, no output tokens burned on prose it doesn't need to generate.
The model runs on Cloudflare's own GPUs under the name @cf/cloudflare/clairic, with a 65,536-token context window and room for up to four images per request. Cloudflare hasn't published parameter counts or a standard accuracy benchmark, but it has shared internal numbers comparing extraction speed against a large language model doing the same job.
Clairic also fits into a bigger framework Cloudflare is laying out, built on the "System One" idea borrowed from TypeSafe's Jev, itself named after Daniel Kahneman's fast-thinking concept. Where Jev and Cloudflare's Clef score pre-listed options, Clairic handles answers nobody could list in advance.
Cloudflare has released a preview of Clairic, a new System One extraction model on Workers AI. The Cloudflare Clairic extraction model reads text or images and returns typed values, not generated text. It targets a job developers usually give to LLMs: pulling totals, dates and IDs into JSON.
Why this matters
For teams running extraction pipelines on top of general-purpose LLMs, Clairic is a pointed bet that the job doesn't need a reasoning model at all. One forward pass, typed fields, no generated tokens: that's a different cost and latency profile than asking GPT-4o or Claude to "return JSON" and hoping the schema holds. The p_null and confidence scores matter more than they might look at first glance.
Most LLM extraction failures aren't wrong answers, they're confident-sounding guesses on fields that were never actually in the document. Giving developers an explicit probability that a value is absent, plus character offsets for text spans, turns extraction into something you can audit and threshold against, rather than something you eyeball and hope.
The missing parameter count and the "preview" label are worth sitting with. Cloudflare hasn't told us how big this model is or what it costs to run at scale, and a 65,536-token window with four images per request is a real constraint for anyone extracting from long contracts or multi-page scans. Worth testing against your actual documents before swapping it in for a bigger model.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Box CEO Aaron Levie Says AI Will Drive "Biggest Cybersecurity Upgrade
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives Multimodal AI Model Clef, Fine-Tuned From Qwen, Comes to Ollama's Cloud
From Archives Kling AI 4.0 adds stereo audio, lip sync; holds own but trails on physics
Common Questions Answered
What is the Clairic System One model and how does it differ from general-purpose LLMs for data extraction?
Clairic is a specialized extraction model released by Cloudflare on Workers AI that reads text or images and returns typed values directly, rather than generating prose. Unlike general-purpose LLMs such as GPT-4o or Claude, Clairic completes extraction tasks in a single forward pass without reasoning steps or unnecessary output tokens, making it more efficient for pulling structured data like totals, dates, and IDs from documents.
What specific data extraction tasks is Clairic designed to handle?
Clairic is designed to extract structured data from invoices, scanned forms, and other documents, pulling values such as totals, dates, IDs, and other typed fields directly into JSON format. The model is optimized for tasks that require returning specific data values rather than generating conversational text, making it ideal for automated extraction pipelines.
How does Clairic's cost and latency profile compare to using general-purpose LLMs for extraction?
Clairic offers a significantly different cost and latency profile compared to general-purpose LLMs because it operates in a single forward pass without generating unnecessary tokens or reasoning steps. This efficiency makes it faster and more economical than asking models like GPT-4o or Claude to return JSON, while also providing confidence scores that help identify extraction reliability rather than relying on the model's confident-sounding but potentially incorrect responses.
What advantage do p_null and confidence scores provide in Clairic's extraction process?
The p_null and confidence scores in Clairic help developers understand the reliability and certainty of extracted values, which is particularly important since most LLM extraction failures stem from confident-sounding incorrect answers rather than obvious errors. These scores enable better error handling and validation in extraction pipelines by indicating when the model is uncertain about a value.
Why is Clairic available as a preview on Cloudflare Workers AI?
Cloudflare released Clairic as a preview on Workers AI to demonstrate that specialized extraction tasks don't require the reasoning capabilities of general-purpose reasoning models. The preview positioning allows developers to test whether a dedicated extraction model better serves their data extraction needs compared to traditional LLM-based approaches.
Further Reading
- Introducing Clairic, a new System One class of extraction models - Cloudflare Blog
- Introducing Clairic: an extraction model for typed values, now on Workers AI - Cloudflare Developers
- Clairic by Cloudflare: extraction model specs, pricing and capabilities - System One Models
