Cloudflare Clairic speeds up data extraction from documents

Cloudflare unveiled Clairic, a specialized AI model to extract data from documents and images faster than an LLM. The tool is in preview on Workers AI, with open weights planned later.

Cloudflare has introduced Clairic, a specialized AI model that extracts data from documents and images faster and at a lower cost than a large language model, according to the company. Available in preview on Workers AI, it mainly targets repetitive tasks such as extracting an order number, an invoice total, or a date. This is important for companies automating customer support, finance, or other document workflows. And for teams tired of reviewing fields one by one, the gain can quickly become noticeable.

A model designed to read, not to chat

Clairic belongs to what Cloudflare calls “System One” models: specialized models designed to perform a specific task, without generating text token by token like a classic LLM. The user specifies the fields sought, then the system returns the values found in the correct format, with a confidence score. If data is missing, Clairic can return null rather than inventing an answer.

This approach places it in a very different logic from generalist assistants. Instead of writing, Clairic extracts, calculates, estimates, and can even produce a simple forecast. Cloudflare also states that the model supports images, allowing it to process photographed receipts or other unstructured documents.

Cloudflare pushes a chain of models rather than a single large model

The company does not present Clairic as a replacement for LLMs, but as a component to insert into a larger chain. In its customer service example, Clairic isolates the order number, amount, and type of damage from a complaint email. Another model, Clef, then applies business rules to decide between refund and repair, then GLM-5.3 drafts the final response.

According to Cloudflare, this chain takes 19.3 seconds. The scenario with GLM-5.3 alone takes 61.7 seconds for the three steps and costs almost twice as much. The company also claims a 69% reduction in processing time and a cost reduced to 6,032 dollars per million customer claims, compared to 11,740 dollars with GLM-5.3 alone. These figures come from internal tests and have not been independently verified [to be verified].

Gains, but no magic

Clairic is not supposed to do everything. Cloudflare acknowledges that the model still struggles with multi-step calculations dependent on each other and recommends entrusting such operations to classic code. The same applies to some forecasts: in a test of daily traffic at TSA checkpoints in the United States over 21 days, the model’s forecasts deviated from actual counts by 3.3% on average, and actual volumes remained within the predicted range 17 days out of 21.

The model features a context window of 65,536 tokens and a starting rate of 0.10 dollars per million input tokens. Since Clairic does not generate output tokens, billing applies only to input. Cloudflare plans to release its open weights later, with no announced date. Until then, the message is quite clear: for many administrative tasks, a smaller model may suffice. The large LLM remains useful when real writing is needed. The rest of the time, it sometimes chats a bit too much.

Key points

  • Clairic is available in preview on Workers AI.
  • Cloudflare announces 19.3 seconds versus 61.7 seconds.
  • Costs drop from 11,740 to 6,032 dollars.
  • Over 21 days, the average error reaches 3.3%.
  • Cloudflare promises open weights later.
  • Clef and GLM-5.3 complete the processing chain.

By the numbers

  • 19.3 seconds — Clairic + Clef + GLM-5.3 workflow, according to Cloudflare.
  • 61.7 seconds — workflow with GLM-5.3 alone, according to Cloudflare.
  • 69% — processing time reduction claimed by Cloudflare.
  • 6,032 dollars — cost per million customer claims, Cloudflare internal test.
  • 65,536 tokens — context window announced for Clairic.

To read