TL;DR: Cloudflare released Clairic on Oct 19, 2026, a model built only for pulling values out of messy input: invoice numbers, totals, dates, line items, and estimates with a range. Instead of writing text, it returns each field as a typed value with a confidence score, or null when the value isn’t there. It’s a preview on Workers AI as @cf/cloudflare/clairic, priced at $0.10 per million input tokens with no charge for output. If you pay an LLM to turn emails, invoices or receipts into JSON, it’s worth a test.

What Cloudflare launched

Cloudflare calls Clairic a “System One” extraction model. It answers in one pass, without reasoning or generating text. You send it the input (text, plus up to four images) and a list of named fields, each with a type such as text, number, date, duration, list, estimate or forecast. You get one answer back per field.

According to Cloudflare’s model docs:

  • Price: $0.10 per million input tokens. Cloudflare says it charges only for input because the model generates no output tokens.
  • Context window: 65,536 tokens. There’s no limit on the number of fields; fields that don’t fit in one 8,192-token pass are split across parallel passes.
  • Images: up to four PNG, JPEG or WebP images per request (4 MiB each, 8 MiB total). Remote image URLs aren’t accepted, so images have to be sent embedded in the request.
  • Status: preview. Cloudflare says it plans to release the model with open weights “soon,” with no date given.

Each answer includes a confidence score and p_null, the probability that the value isn’t in the input. Text answers come back as the exact span of the input, with character offsets, so you can highlight where a value came from.

Why it matters for small teams

Pulling a total, a due date or an order number out of an email is one of the most common jobs automations hand to an LLM. Cloudflare’s pitch is that a typed model can’t return broken JSON or a misformatted number, and that a missing field comes back as null rather than a plausible guess.

In Cloudflare’s own test of a damaged-product claim, a chain of Clairic, Cloudflare’s Clef decision model and an LLM for the reply took 19.3 seconds end to end, against 61.7 seconds when one LLM (GLM-5.3) did every step. Cloudflare puts the cost at $6,032 per million claims against $11,740, about 49% less. Those are Cloudflare’s numbers from one workflow, not an independent benchmark.

Cloudflare is open about the limits of this early checkpoint:

  • One- and two-step math (totals, differences, date arithmetic) seems reliable, but longer chains and adding up many table rows aren’t. Cloudflare’s advice: “have Clairic predict the value and let code do the math.”
  • Estimates need evidence in the input, such as comparable sales or history.
  • Forecasting is “basic” and isn’t meant to replace a real time-series model.
  • “The architecture makes every answer well formed, not necessarily correct.” Use the confidence scores to decide what to automate and what a person should check.

What to do

  1. Pick one extraction step you already run. Invoice intake, order emails, receipts or support tickets are good candidates. Note what your current LLM call costs and how often it fails.
  2. Run the same inputs through Clairic. Call @cf/cloudflare/clairic from a Worker or the Workers AI REST API, with one typed field for each value you need.
  3. Set a confidence cutoff. Send low-confidence or null answers to a person instead of straight into your accounting or CRM tool.
  4. Keep the math in code. Let Clairic read the numbers and let your workflow add them up.
  5. Don’t build anything critical on it yet. It’s a preview, and Cloudflare’s model terms say it may stop providing the model.

If your agents already act on tool outputs, our guide on stopping AI agents from trusting the wrong tool outputs covers how to add checks like these. For a lighter forecasting option that runs on a CPU, see our coverage of Mistral’s Timestral models.

Sources & method

Sources

Method. Last checked October 20, 2026. Price, context window, image limits and request rules come from Cloudflare’s model docs; the speed and cost comparison, limits and open-weights plan come from Cloudflare’s blog post. We haven’t tested Clairic ourselves. No affiliate links on this page. How we work.

Spotted an error? See our corrections policy and tell us.