Cloudflare's Clairic Is an AI That Doesn't Write — It Just Extracts, 70% Faster

Cloudflare's new extraction AI model doesn't generate text at all

Cloudflare announced Clairic on October 19: an extraction AI model that pulls numbers, dates, names, and other hard data directly from text and images — without generating a single word of prose. The company claims it cuts processing time by 70% compared to large language models, returning answers in under one second. It is now in preview on Cloudflare's Workers AI inference platform.

The pitch is a direct attack on how everyone currently uses LLMs for document processing. Today, to pull invoice totals or contract dates from messy PDFs, you make a chatbot read the document and write back an answer. That takes time, the model can invent numbers that were never in the file, and you still need a validation step to check the format of whatever it wrote. Clairic processes the input once and returns the answer directly. No prose, no formatting surprises.

Billing matches the philosophy: you pay for input tokens only, at $0.10 per million tokens. No output token charges because there is no output generation.

System One: the decision-model wave hits extraction

Cloudflare calls Clairic the first in a new category it dubs "System One extraction models" — a nod to psychologist Daniel Kahneman's fast, intuitive thinking. The framing is deliberate: since TypeSafe unveiled its Jev decision model in September, a whole sub-industry has formed around models that choose rather than write. Cloudflare itself shipped Clef on October 1, plus Clef-omni for audio and video, as open-weight rivals to Jev.

Jev and Clef select answers from pre-prepared options. Clairic pushes the idea further: instead of choosing between options, it extracts specific values and can forecast numbers. Cloudflare claims it is the first to extend the "System One" concept to information extraction and numerical forecasting. Whether the label sticks or not, the technical direction is clear — for classification and extraction workloads, generation is expensive overhead.

That connects to a broader shift we've been tracking. TypeSafe's Jev went from $200M to a $7.5B valuation in 24 days precisely because enterprises are hungry for models that give deterministic answers instead of chatty ones. Clairic is Cloudflare's argument that the same logic applies to every document pipeline in the company.

What it actually changes for builders

The obvious use cases are the boring ones that cost real money: invoices in ten different formats, customer support emails with order numbers buried in rants, forms photographed at odd angles. LLMs handle these today but wastefully — you pay for generated tokens, wait for the answer to be typed out, then parse and validate it. An extraction model that answers in under a second, billed on input only, changes the unit economics of high-volume document processing.

The caveats matter. Clairic is in preview, so reliability claims are Cloudflare's own until independent testing lands. Extraction is also a narrower problem than general reasoning: if your pipeline needs the model to reason across documents or draft responses, you still need an LLM. And the 70% figure is a company benchmark against LLM baselines — real-world speedups will depend on document complexity and the accuracy bar you set.

Still, watch the pattern: first decision models replaced prose with probabilities, now extraction models are replacing prose with answers. The chatbot era of AI infrastructure had a great run, but the money is moving toward models that do one thing fast. Cloudflare's recent moves — from buying Deno to extend its runtime reach to shipping three in-house model lines in a month — suggest it wants to own that infrastructure layer. With GPT-6 rolling out to every ChatGPT user on the consumer side, the enterprise side is clearly fragmenting into specialized models. Clairic is the latest — and probably not the last — entry in the post-chatbot model zoo.