What happened
Cloudflare’s November 18, 2026 announcement says its Workers AI Superintelligence team has redesigned Clairic since the model’s preview release and is introducing Clairic-flash. The updated Clairic is the 27-billion-parameter model; flash is a lighter 9-billion-parameter version. Both are extraction models, intended to return structured fields rather than general-purpose free-form answers. Cloudflare says both are hosted on Workers AI and that full model weights are published on Hugging Face.
The redesign changes how ordinary text and number fields are generated. Instead of producing all output in a single forward pass, Clairic uses parallel slot autoregressive generation: each field uses the tokens it has already generated to choose its next token, then ends at its own marker. Cloudflare says fields advance together in the serving batch, preserving fast per-request latency. The company also trained special heads for estimates, forecasts, durations, calculated dates, and explicit arithmetic. Its typed API retains confidence scores, character spans, and probabilities, and the API schema is unchanged from preview.
The release extends document handling to PDFs. A request can include up to four images or PDF pages in total, and line items can be returned as rows that keep description, quantity, and price together. Cloudflare says the revised image-reading system recorded zero errors in its internal benchmark, compared with 207 errors before; those are company-run results, not a public independent evaluation. Its broader table reports improved counts over preview on several internal extraction tests, including synthetic receipt fields and tracked customer assertions. The announcement explicitly calls the evaluations directional and self-reported, and says no public extraction benchmarks exist yet.
Cloudflare lists Clairic at $0.24 per million input tokens and Clairic-flash at $0.10 per million input tokens. It says it charges only for input tokens, absorbing the cost of generated output. Until December 1, 2026, both models are promotional priced at $0.10 per million input tokens. Existing applications using the preview Clairic endpoint are being upgraded in place; customers can move cost-sensitive workloads to the flash endpoint. The company reports self-hosted HTTP response latency below one second in its tests, with results varying by workload and model.
What the archive adds
New Runtime covered Clairic’s initial release on October 22, 2026, describing its typed extraction interface. This is a new event on that story: the model has been redesigned, extended to PDFs, made open-weight, and joined by a smaller variant. The distinction matters because this is more than a restatement of the preview announcement.
Trend signal
This release supports the idea that model answers are becoming bounded task interfaces: Clairic returns typed fields with confidence scores, spans, and probabilities rather than only prose. New Runtime tracks that shift in Model answers become task interfaces. Its open weights also fit the assumption that some teams will adopt open models through managed GPU services rather than only hosted token APIs, described in New Runtime’s analysis of managed open-model adoption.
Why now
Cloudflare says customer use cases and feedback after the preview included extracting receipt data from images and reading values from support tickets. It redesigned the model to focus on unstructured-data extraction, and now presents the updated model alongside flash and open weights. The promotional price ends December 1, 2026, creating a defined period to evaluate the larger model at flash’s listed price.
Connections to other events
New Runtime’s October 10, 2026 coverage of Cloudflare’s Clef-omni and Clef models noted Cloudflare’s release of multimodal models. Clairic extends the company’s model lineup in a different direction: focused extraction with a typed output interface, rather than multimodal general-purpose input. Together, the releases show Cloudflare presenting multiple purpose-built model families for Workers AI and open-weight experimentation.
Why it matters
Teams often spend large-model tokens extracting a small set of facts from documents, even when the task is repetitive and the desired answer has a fixed schema. A specialized extractor could reduce latency and cost, but it must preserve source grounding and handle messy inputs well enough to avoid creating more review work than it removes. Clairic’s confidence scores and character spans provide useful signals for downstream validation, while its PDF and multi-page support address a common document workflow.
The pricing makes a practical comparison possible: until December 1, teams can evaluate the 27B Clairic at the same input-token price as the 9B flash model, then compare accuracy and latency against flash. Open weights also let teams test self-hosting or adaptation where data handling or deployment control matters. Those options are not proof of production performance: Cloudflare’s accuracy claims come from internal evaluations, so teams should test representative documents and retain human checks for consequential fields.
What happens next
- Cloudflare will end the Clairic promotional price and restore its $0.24-per-million-input-token rate. (3 months, 95%). Watch for: Workers AI pricing after December 1, 2026.
Limits and uncertainty
Cloudflare’s extraction evaluations are internal and self-reported; the company says public benchmarks are not yet available. The announcement does not establish how the models perform across independent datasets, languages, document formats, or production distributions beyond the tests described. Its latency figures come from a self-hosted server and are workload-specific, so they do not independently establish Workers AI response times. The release also does not quantify the operational cost of self-hosting, adaptation, or review of uncertain outputs. Builders should validate accuracy and latency against their own documents before replacing an existing extraction path.
Skill to build now
Build a small evaluation set from real documents before switching an extraction workload. Include varied scans, multi-page PDFs, tables, dates, and arithmetic fields; define expected typed outputs and flag the fields where errors have the highest cost. During the promotional period, run the same requests against Clairic and Clairic-flash, recording correctness, latency, and the usefulness of confidence scores and character spans. Keep the existing workflow as a fallback until those checks pass.
Reader takeaway
Clairic’s update makes document extraction more concrete: PDFs and images, typed fields, source spans, and open weights arrive with a smaller companion model. The strongest reason to test it is not Cloudflare’s internal benchmark alone, but the chance to compare the two sizes on your own documents while the larger model is temporarily priced like flash.