Perfect scores, on tests the vendor wrote

The strongest numbers in Cloudflare's announcement are perfect ones. On synthetic receipt fields, both new models score 576 out of 576 — fields counted across Cloudflare's synthetic receipts. On tracked customer assertions, both score 1,263 out of 1,263, assertions counted across 139 request shapes. On fresh text fields, the 27B Clairic gets 335 of 336 fields right. None of those figures is a document-level accuracy rate, and none was measured on a buyer's own invoices.

Cloudflare, to its credit, prints the qualification in the same post. There are no public extraction benchmarks yet. The evaluations are internal, the competitor rows were run by Cloudflare itself, and the company instructs readers to take them as directional guidance because they are self-reported. That sentence does more work than any score in the table. It means the 576/576 tells you the models handle Cloudflare's synthetic receipts, written to Cloudflare's field definitions, under Cloudflare's scoring — and nothing yet about documents the vendor did not design the test around.

What actually shipped

Underneath the scores there is a concrete release. Clairic has been redesigned since its October preview: for ordinary text and number fields it now uses parallel slot autoregressive generation instead of a single forward pass — each field choosing its next token from the tokens it has already written, then stopping at its own end marker, in Cloudflare's description. Trained heads handle the field types plain extraction tends to fumble, including estimates, forecasts, durations, calculated dates and arithmetic. PDF reading is new. A single request can take up to four images or PDF pages and return a supplier, its line items as intact rows, and a total that sits on a different page. Cloudflare says the typed interface is intact with every request — a confidence score, a character span and probabilities with each answer — which is the part of the design that matters in production: a field the model is unsure about can be routed to a person instead of silently entering a ledger.

Clairic-flash is the practical half of the release. At 9 billion parameters it gives up seven fresh text fields out of 336 to the larger model in Cloudflare's table, matches it on the customer assertions and the synthetic receipts, and on fresh text fields in Cloudflare's tests answers at a 155 ms median against the larger model's 358 ms. Both models are on Hugging Face as open weights under Apache 2.0, so a team can run the evaluation Cloudflare cannot run for it, on its own documents, this week.

The endpoint swap is the part to act on

One operational detail deserves more attention than the launch framing gives it. The Workers AI endpoint that preview users already call, @cf/cloudflare/clairic, used to serve the 9B model. It now serves the 27B model, under an unchanged endpoint name; the 9B model has a new endpoint, @cf/cloudflare/clairic-flash. Output, latency and price can all change for an existing integration: both models cost $0.10 per million input tokens until December 1, 2026, and Clairic is listed at $0.24 after the promotion, while Clairic-flash stays at $0.10. Anyone running the preview in production has effectively been migrated to a different model. The published tests do not establish how output or latency will change for a particular deployment, so re-run your own acceptance tests on your own documents before relying on either endpoint, and before the price changes, not after an invoice or an error report raises the question.

Sources

  1. [1] Cloudflare, "An updated Clairic and Clairic-flash are now open-weight, from the new Workers AI Superintelligence team" — read first-hand: the evaluation and latency tables, the architecture, pricing, endpoint changes and the no-public-benchmark caveat.Read source
  2. [2] Cloudflare, Clairic and Clairic-flash model cards on Hugging Face — Apache 2.0 license and repository contents verified first-hand.Read source
  3. [3] Unite.AI, "Cloudflare Releases Open-Weight Clairic Models for Data Extraction" — secondary; the October 19, 2026 preview date.Read source