Cloudflare Releases Clairic, a System One Extraction Model on Workers AI

Cloudflare has put Clairic into preview on Workers AI. It is an extraction model: it reads text or up to four images and returns typed values such as totals, dates, durations, estimates and forecasts, instead of writing prose. Every requested field gets answered in a single forward pass, with no reasoning tokens and no output bill. Pricing is $0.10 per million input tokens.

That matters because a huge share of what businesses actually do with large language models is exactly this chore: read an invoice, a claim or an email and fill in a form. Today many teams pay a general-purpose reasoning model to do it, wait seconds for the answer, and then hope it formats the JSON correctly.

In Cloudflare's own damaged-product claims test, the extraction step took 0.46 seconds with Clairic versus 16.1 seconds with GLM-5.3, and the full pipeline cost per million claims fell 49%. Those are vendor numbers, but the direction is the story.

We like where this is going. The industry spent two years using the most expensive tool in the box for every job, and the bill showed up in latency, cost and flaky outputs. Cloudflare's framing is a sensible division of labor: extraction models turn raw input into values, decision models like its own Clef turn values into choices, and an LLM only gets involved at the end to write words a human will read. That is how mature software gets built, with specialized parts and clear contracts between them.

The design details point in a good direction too. Each answer carries a confidence score and a separate probability that the value simply is not there. In Cloudflare's examples, a missing tracking number came back empty rather than as a plausible fake, and a receipt photo returned nothing for a discount that did not exist. A model that is built to say it does not know is worth more in an accounts-payable workflow than a smarter one that guesses with confidence. Repeatability matters as well: Clairic gave the same repair estimate on every run, while the LLM drifted between $250 and $220.

Now the caution. There is no standard accuracy benchmark published yet, the parameter count is undisclosed, and the weights are only promised. Cloudflare itself says multi-step arithmetic is unreliable, and four of its 21 next-day travel forecasts landed outside the stated 80% range, three of them around a holiday weekend. That is honest disclosure, and it is also a warning: these numbers will look authoritative because they arrive as clean typed fields with confidence scores attached.

That is the harm we would flag if this pattern scales. When an LLM produces a wrong number in a paragraph, a reader might squint at it. When an extraction model writes a wrong total straight into a database, nobody squints. Confidence scores only help if teams actually set thresholds and route low-confidence fields to people. There is also a quieter cost trap: very wide schemas are billed as input, and in the docs example the field definitions outweighed the invoice itself.

Our read: this is useful plumbing, and cheaper, faster, more predictable extraction is a real win for small teams. I'm watching for independent accuracy numbers and for the open weights, because until then the trust in this layer rests on one company's tests.

Clairic follows TypeSafe's Jev and Cloudflare's Clef, both decision models that pick among options a developer supplies. Clairic covers the answers that cannot be listed in advance, like an invoice total or a repair cost. It handles a 65,536-token context, splits large schemas into parallel passes, and is available through the Workers AI binding, the REST API and AI Gateway.