Cloudflare released Clairic, an extraction model that generates no tokens

An extraction job that calls an LLM for every receipt pays by the generated word, even when every field in the answer was already in the input.

Cloudflare released Clairic, a model it describes as "a System One extraction model" that reads unstructured input and returns typed values in a single forward pass. It generates no output tokens. The model is available in preview on Workers AI as @cf/cloudflare/clairic, and Cloudflare says it intends to release the model with open weights.

What it returns

Clairic emits four kinds of value, each described in Cloudflare's post. An extracted value is a text span from the input, with its character offsets and a confidence score. A computed value is a number, date, duration or list the model reads from the input or works out from it. An estimated value, for a field the input does not state outright, comes back as "a distribution inside the bounds you declare, on the grid you specify, with a median and an 80% interval," plus every percentile on request. A forecasted value is a future step of a series, each with its own 80% interval.

Cloudflare says the model "produces well formed, not necessarily correct answers," and lists three limits: multi-step arithmetic is unreliable and should be moved to code, estimates need supporting evidence in the input, and the forecasting gives a quick signal while a real time series model stays the right tool for serious forecasting work.

The pipeline comparison

The post walks through a receipt test and a damaged-product-claims pipeline. On the receipt, Cloudflare reports confidence scores from 0.776 to 0.9999 across the extracted fields. On the claims pipeline, which chains Clairic with a classifier and an LLM for the final reasoning step, Cloudflare says the end-to-end run took 19.3 seconds against 61.7 seconds for an all-LLM pipeline, and cost $6,032 per million claims against $11,740 for the all-LLM baseline. The 49 percent cost saving and the 69 percent latency saving are Cloudflare's own figures on a pipeline it built for the comparison.

A separate forecasting test used 21 days of TSA checkpoint data. Cloudflare says "Clairic's forecasts were off by 3.3% on average, and 17 of the 21 actual counts landed inside its 80% interval."

The baselines

Cloudflare benchmarks Clairic against GLM-5.3, Gemma 4 and Qwen in the post. On the damaged-claims task, Cloudflare says Gemma 4 and Qwen "failed at this task, as they were unable to deduce values like repair_cost from the inputs." Those are Cloudflare's own tests against models it did not train, so read them as vendor benchmarks.

The model class matters here more than the specific numbers on this one pipeline. An extraction step that runs through a chat model today is billed on the words it generates, and for a filled form those words are the field values retyped. A model that generates nothing removes that line from the bill. Whether the accuracy is good enough on a team's own documents is a question only that team's own test set can answer.

Source

• Cloudflare, Introducing Clairic, a new System One class of extraction models