Cloudflare introduced Clairic on October 19, 2026, releasing a preview of what it describes as the primary mannequin in a brand new class of System One extraction fashions, hosted on Staff AI with an open-weights launch deliberate for later.
The mannequin is constructed to drag typed values out of unstructured inputs, extracting, computing, estimating, and forecasting slightly than producing textual content. Cloudflare says Clairic does this at greater accuracy, sooner latency, and decrease price than conventional massive language fashions.
Availability and Pricing on Staff AI
Clairic is accessible in preview as @cf/cloudflare/clairic, callable from a Cloudflare Employee with env.AI.run() or via the REST API. The official mannequin documentation lists a 65,536-token context window, imaginative and prescient assist, and unit pricing of $0.10 per million enter tokens.
As a result of the mannequin generates no output tokens, billing applies to enter solely. Cloudflare mentioned it intends to launch the weights overtly after additional checkpoint work, with out giving a date.
How the Extraction Mannequin Works
A request pairs a state, the textual content to learn plus as much as 4 non-compulsory photographs, with a map of typed fields. Supported area sorts are textual content, quantity, date, time, datetime, length, checklist, estimate, and forecast, every with non-compulsory plain-language directions. The mannequin solutions each area in a single parallel ahead move, with no reasoning step and no generated tokens, and Cloudflare says it usually responds in beneath a second. Fields that don’t match beside the enter in a single 8,192-token move are break up throughout parallel passes that every see the entire enter, then merged. The documentation limits embedded photographs to 4 PNG, JPEG, or WebP recordsdata of as much as 4 MiB every.
Each scalar reply carries p_null, the likelihood that the worth is absent from the enter, and a confidence rating. Estimate fields return a distribution inside caller-declared bounds, with a median and an 80% interval from p10 to p90 and all 99 percentiles on request. A area the enter doesn’t decide returns null, with a guess area exhibiting what the mannequin would in any other case have picked, slightly than a fabricated worth.
Cloudflare locations Clairic within the System One class alongside the choice fashions that preceded it. TypeSafe launched its Jev resolution mannequin a few month earlier than the announcement, and Cloudflare launched its Clef and Clef-flash resolution fashions on October 1, 2026, open-sourcing them on Hugging Face beneath an Apache 2.0 license; Clef was post-trained on Qwen base fashions and is Jev-API suitable. The place resolution fashions classify an enter amongst listed choices and enormous language fashions generate open-ended textual content, Clairic returns typed values instantly: spans, numbers, dates, lists, and distributions. The corporate describes the three classes as complementary levels, with extraction feeding selections and selections feeding the ultimate generated reply.
Documented Examples and Reported Exams
In Cloudflare’s support-email instance, the mannequin computed a $701.10 whole from a said $649.00 value plus $52.10 tax in the identical move, derived 11 days between the September 28, 2026 supply and the October 9, 2026 message date, listed cracked display and dented nook as precise spans of the enter, and returned null for a monitoring quantity the e-mail by no means contained. In a receipt-photograph instance, Clairic extracted the order quantity, buy datetime, amount, unit value, subtotal, gross sales tax, and tip with confidence scores, and returned null for a reduction the receipt didn’t present.
For estimation, the announcement exhibits Clairic pricing a house from comparable-sales textual content: the mannequin returned a price of $418,000 with a median of $415,000 and an 80% interval of $370,000 to $470,000 inside caller-set bounds. For forecasting, Cloudflare reviews a 21-day check through which Clairic predicted next-day TSA airport checkpoint screening counts from the general public collection. The corporate says the forecasts had been off by 3.3% on common and that 17 of the 21 precise counts landed contained in the mannequin’s 80% interval, with three of the 4 misses falling on the Columbus Day weekend. For October 14, 2026, Cloudflare reported an precise rely of two,419,474 towards a predicted median of two,419,283.
Cloudflare additionally reviews a pipeline comparability that pairs Clairic with Clef and the GLM-5.3 language mannequin towards an all-GLM-5.3 model of the identical damaged-product claims workflow. In keeping with the corporate, the cascade completed in 19.3 seconds towards 61.7 seconds, 69% much less time or about 3.2 instances sooner, and value $6,032 per million claims towards $11,740, or 49% much less. Clairic and Clef dealt with extraction and the choice in beneath a second mixed, the place GLM-5.3 took 38 seconds for a similar two steps, and Cloudflare says Gemma 4 and Qwen couldn’t derive values comparable to restore price and failed the duty within the all-LLM model. Cloudflare constructed the workflow right into a dwell demo.
Preview Limitations and Open-Weights Plans
Cloudflare calls the present endpoint an early checkpoint and lists a number of limitations. One- and two-step arithmetic comparable to totals, variations, conversions, and date arithmetic appear dependable, it mentioned, however longer chains of dependent steps and aggregation throughout many desk rows are exhausting for a one-pass mannequin with out a scratchpad; it advises customers to depend on Clairic for the prediction and deal with the arithmetic in code. Estimates want proof within the enter, calibration is measured slightly than assured and intervals can run slender on unfamiliar knowledge, and forecasting is primary slightly than an alternative choice to devoted time-series fashions.
The corporate advises utilizing confidence scores, p_null, and the intervals to resolve what to automate and what to route for extra evaluation, as a result of the mannequin’s solutions are assured to be properly shaped, to not be proper. Cloudflare mentioned it should maintain engaged on the arithmetic, estimation, and forecasting sides of the mannequin with every checkpoint, and that it plans to launch Clairic open-weights afterward. The title, it added, performs on clerical work, readability, and clairvoyance, and doubles as shorthand for Cloudflare’s AI regression mannequin.
