An invoice total can be a valid number and still be wrong. Keep format checks and evidence checks separate when testing extraction.
Clairic returns typed values from supplied material
Cloudflare’s October 19 release note introduces Clairic on Workers AI. The model page describes a preview with vision support and a 65,536-token context window.
The interface supports text spans, numbers, dates, lists, estimates and forecasts. Answers include signals such as confidence and the probability that the input does not determine a value. Cloudflare lists $0.10 per million input tokens without an output-token charge.
Check the field against its source evidence
Start with fields whose correct answers can be independently checked: an invoice number, the printed currency and a due date derived from explicit terms. Include documents where a requested field is absent.
For an extracted text span, verify that its location points to the intended passage. For a computed number, calculate the result separately. A schema can enforce a type without proving that the value is supported.
Keep preview limitations in the test
Cloudflare says the early checkpoint has limits in long calculation chains and aggregations across many table rows. It recommends extracting the inputs and doing that arithmetic in code.
Its release note also describes basic forecasting and uncertainty intervals that may become too narrow on unfamiliar data. Open weights are planned; this review did not verify their release.
For dates that depend on today, supply an explicit reference date. Otherwise, a well-formed date can hide the wrong year or point in time.
Choose the review threshold from your own documents
I would build a small labeled collection with clean invoices, ambiguous totals, missing purchase-order numbers, and unusual layouts. Keep a separate evaluation set when choosing a review threshold.
- Check exact matches for identifiers and dates.
- Recompute totals using ordinary application code.
- Measure errors among fields accepted automatically.
- Inspect how often missing fields produce a value.
- Count review time and cost per accepted document.
Confidence values are model outputs. Their usefulness depends on how they behave on your material. We have not run a hands-on accuracy or speed comparison for this article.
The updated Clef guide covers a model for defined decision options. Our Timestral evaluation guide explains a different forecasting workflow.
Read the sources and plan your next step
- Clairic release, API fields and preview caveats
- Model context, access and published price
- Cloudflare’s extraction-model explanation
Which missing or ambiguous field should your first test deliberately include?
Checked October 20, 2026. This is source-based reporting and analysis; proposed checks are not hands-on test results.
