Cloudflare’s Clairic aims to extract answers without making an LLM write them

Cloudflare has introduced Clairic, a preview model for extracting, calculating and estimating values from text and images. Its practical pitch is typed results, source spans and confidence scores in a single pass, rather than a language model generating answers that then need to be parsed and checked.

Cloudflare Watch analysis

What happened

Cloudflare says Clairic is available in preview through Workers AI as cf/cloudflare/clairic, with open weights planned for a later release. The company describes it as the first in a new class of “System One” extraction models, designed for tasks such as finding invoice details, calculating totals, estimating a repair cost or forecasting a time series.

The Clairic announcement says outputs can include exact text spans with character offsets, typed dates and numbers, estimates with an 80% interval, and a confidence score. Users specify the fields they want, including each field’s type and optional instructions. The model can also return null when it cannot find an answer. Cloudflare says the model accepts up to four optional images alongside text.

Why it matters

Data extraction is a common reason to put an LLM into a workflow, but generated text can be awkward to validate: a date might be malformed, a field might be missing, or a plausible answer might not be traceable to its source. Clairic’s combination of typed output and text spans is aimed at making those results easier to inspect and use in software.

Cloudflare says Clairic makes one parallel pass, generates no output tokens and typically responds in under a second. It also says multiple fields cost about the same as a smaller set, unless they need to be split across passes. Those are the company’s product claims; the announcement does not provide comparative benchmark results. Still, the useful distinction is clear: extraction and estimation are being offered as a specialised alternative to asking a general-purpose chatbot to do everything.

Our read

This is a promisingly specific tool for developers dealing with messy documents, not a replacement for language models across the board. The source spans, null handling and typed fields are the details to try first. They could make errors easier to spot, which is rather more useful than merely making them arrive quickly.

Cloudflare says an open-weights release is coming, but gives no date in the announcement. For now, developers can assess the preview on their own inputs and compare its accuracy, latency and cost with the extraction pipeline they already use.

What to watch

• Whether Cloudflare publishes comparative accuracy, latency and cost results.
• How reliably confidence scores and null handling flag uncertain or missing values.
• When the promised open-weights release arrives, and what its licence permits.

Discussion spark: Would you trust a specialised model’s typed fields and confidence scores for document processing, or should every extracted value still be checked against its source?

not affiliated with or endorsed by Cloudflare