Pulling a few key facts from a document can be a different job from writing a fluent response about it. Cloudflare’s Clairic is designed for that narrower task: users specify the fields they need, and the model returns matching values in a defined format, along with a confidence score.
Purpose-built for extracting fields
A developer might ask for an order number or an invoice total. Clairic looks for those requested values and returns them directly. If a field is absent, it responds with null, making missing information explicit rather than filling the gap with generated text.
That approach differs from large language models (LLMs), which generate text token by token. Clairic handles every requested field in a single pass. For document-processing work, this design means the task can be framed as extracting specific data rather than asking a general-purpose model to interpret a document and compose an answer.
Cloudflare says Clairic can also perform simple calculations, estimates, and forecasts. That extends its role beyond copying values from a document, but the company cautions that more involved calculation chains are not yet reliable enough to hand over to the model.
One model in a larger workflow
Cloudflare describes Clairic as a component that can work alongside LLMs and other “System One” models, including Jev. In this setup, different models handle distinct stages of a task: one gathers facts, another applies a decision process, and a language model can prepare the final written response.
The company’s complaint-email example shows how that division could work. Clairic extracts the order number, amount, and damage from the message, then estimates the repair cost. The decision model Clef applies company rules to select a refund or a repair, while GLM-5.3 writes the customer reply.
Separating those steps gives each part of the workflow a defined responsibility. The extracted information supplies inputs for the decision; the company rules guide the choice; and the writing model turns the result into a response. In Cloudflare’s example, the workflow combines extraction, decision-making, and drafting rather than asking one model to do all three.
Cloudflare’s speed and cost comparison
Cloudflare says its three-step chain takes 19.3 seconds. It compares that with GLM-5.3 handling all three steps on its own, which takes 61.7 seconds and costs almost twice as much.
Those figures are Cloudflare’s comparison for the described example. They illustrate the company’s case for assigning extraction and other structured tasks to a model designed for them, while using a language model where generated prose is needed. They do not establish that the same time or cost difference will apply to every document or workflow.
Availability and current limits
Clairic is available in preview on Workers AI, and Cloudflare says open weights are expected to follow soon. The preview status and planned release are relevant for developers considering how the model might fit into a document-processing system.
For now, Cloudflare says Clairic cannot reliably handle longer chains of calculations. Developers should run those calculations in their own code. That boundary matters when designing a workflow: Clairic may extract fields and handle simple estimates, but a longer sequence of calculations should not depend on it alone.
The model’s practical role is therefore specific. It can return requested document values in a predictable structure and contribute to a larger chain of AI tools. Cloudflare’s example suggests potential savings in time and cost, while its stated calculation limitation points to a need to keep more complex arithmetic in developer-managed code.