Summary: Cloudflare’s Workers AI superintelligence team shipped updated Clairic and a new Clairic-flash on November 18, 2026, with full weights on Hugging Face and both hosted on Workers AI. The 9B flash variant posts p50 latency of 155 ms on fresh text fields versus the 27B model’s 358 ms, and both scored 1,263/1,263 on Cloudflare’s tracked-customer extraction eval — numbers the company itself flags as self-reported. For builders parsing receipts, invoices, leases, or PDFs, the promo price makes this cheap to try before December 1, 2026.
What changed since the preview
Cloudflare redesigned Clairic’s architecture rather than iterating on it. Ordinary text and number extraction fields now use parallel slot autoregressive generation instead of a single forward pass: each field reads the tokens it has already written to pick its next token, then stops at its own end marker, with all fields advancing together in the serving batch. The stated trade is a small latency cost for higher extraction accuracy.
The model also gains trained heads for computation-heavy cases — estimates, forecasts, durations, calculated dates, and explicit arithmetic. The typed API stays the same, keeping confidence scores, character spans, and probabilities on every request.
Document input widened too. When Clairic first launched in preview, it read a photo of a receipt; now the family reads PDFs, up to four images or PDF pages per request, with line items returned as rows that keep description, quantity, and price together. Cloudflare reports image reading errors dropping from 207 to 0 on its internal benchmarks.
The eval numbers, and who produced them
The headline result: on tracked customer assertions across 139 request shapes, both Clairic and Clairic-flash scored 1,263/1,263, against Clairic-preview at 1,033, Liquid AI’s D1.2 pipelined with Apple Vision OCR at 973, and Scaledown’s Extract at 653. On synthetic receipt fields in text and images, both new models hit 576/576 versus 397 for D1.2 + OCR.
Read these as marketing-adjacent, because that is what they are. Cloudflare states plainly that there are no public extraction benchmarks yet and that its evals are self-reported and should be taken as directional guidance. The comparison to D1.2 and Extract is Cloudflare running competitors through its own harness — useful as a signal, not as a reason to migrate.
Latency on a self-hosted server
The latency figures come from complete HTTP responses against a self-hosted server, including all fields per request:
| Eval set | Requests | Clairic p50/p95 | Clairic-flash p50/p95 |
|---|---|---|---|
| Fresh text fields | 96 | 358 / 736 ms | 155 / 328 ms |
| Fresh image fields | 48 | 453 / 976 ms | 305 / 605 ms |
| Commercial leases | 180 | 691 / 803 ms | 379 / 519 ms |
| Loan pricing | 600 | 395 / 407 ms | 188 / 192 ms |
| Traffic forecasting | 114 | 532 / 585 ms | 243 / 261 ms |
| Expense audit | 60 | 632 / 742 ms | 393 / 512 ms |
Two things stand out. Clairic-flash beats the 27B model on every set, and its p95 on loan pricing — 192 ms across 600 requests — barely moves above its p50, suggesting consistent performance on that workload. But the measurements are self-hosted, not Workers AI, so hosted latency may differ.
Pricing: input tokens only, with a deadline
The 27B Clairic will cost $0.24 per M input tokens; the 9B Clairic-flash stays at $0.10. Both bill only on input tokens — the output generation cost is absorbed by Cloudflare. Until December 1, 2026, a promotional discount holds both at $0.10. Existing apps pointed at the preview endpoint get the upgraded model in place, no code change required.
That deadline is the real news for adoption math. After December 1, the 27B model costs 2.4× the flash variant while the evals show identical accuracy on the tracked customer set — so the honest default is to benchmark flash on your own documents and pay the 27B premium only where it demonstrably wins.
Open weights, licence unstated
Both models are open-weight on Hugging Face, following the pattern of Cloudflare’s earlier Clef release, which drew community experiments and fine-tunes. The post does not state the licence terms on those weights, so anyone planning self-hosting for commercial use should check that before committing — hosted access on Workers AI and self-hosted weights are different commitments with different obligations.
Our take: If your product feeds documents into a general-purpose LLM for extraction, spend an afternoon on this now — swapping an endpoint is reversible and the promo keeps the 27B model at flash pricing until December 1, 2026. The reasoning: extraction-only billing means retries and failed parses cost less than LLM-token pricing, and flash’s sub-200 ms p50 on the loan-pricing set suggests the finished task gets faster, not just cheaper. We assume the self-reported evals hold roughly on real documents; that is the assumption to test, not trust. Ignore this if your workload is open-ended generation or reasoning — this family is purpose-built for extraction, and your existing tools already fit. The verdict flips if a self-test shows flash diverging from the 27B model on your document mix, or if the Hugging Face licence blocks commercial self-hosting and hosted-only latency disappoints.
Sources
View the original post on X
