Cloudflare releases Clairic, its first System One mining model – Unite.AI

Cloudflare announced Clairic on October 19, 2026, releasing a preview of what it describes as the first model in a new class of System One extraction models, hosted on Workers AI with an open-weight release planned for later.

The model is built to extract typed values ​​from unstructured input, extracting, calculating, predicting and predicting rather than generating text. Cloudflare says Clairic does this with higher accuracy, faster latency, and lower cost than traditional large language models.

Availability and pricing on Workers AI

Clairic is available in preview as @cf/cloudflare/clairicrecallable by a Cloudflare Worker with env.AI.run() or via the REST API. The model’s official documentation lists a 65,536 token context window, visual support, and a unit price of $0.10 per million input tokens.

Since the model does not generate output tokens, billing only applies to input. Cloudflare said it plans to openly release the weights after further checkpoint work, without giving a date.

How the extraction model works

A request pairs with statethe text to read plus up to four optional images, with a typed map fields. Supported field types are text, number, date, time, timestamp, duration, list, estimate, and forecast, each with optional plain language statements. The model responds to each field in a single parallel forward pass, with no reasoning steps and no generated tokens, and Cloudflare says it typically responds in less than a second. Fields that don’t fit next to the input in an 8,192-token pass are split into parallel passes where each sees the entire input, then merged. The documentation limits embedded images to four PNG, JPEG, or WebP files of up to 4 MiB each.

Every scalar response leads p_nullthe probability that the value is absent from the input and a confidence score. The prediction fields return a distribution within the bounds declared by the caller, with a median and 80% range from p10 to p90, and all 99 percentiles on request. A field that the input does not determine returns null, with a guess field showing what the model would have otherwise chosen, rather than a fabricated value.

Cloudflare places Clairic in the System One category alongside the decision-making models that came before it. TypeSafe introduced its Jev decision model about a month before the announcement, and Cloudflare released its Clef and Clef-flash decision models on October 1, 2026, making them open source on Hugging Face under an Apache 2.0 license; Clef has been post-trained on the Qwen base models and is compatible with Jev-API. Where decision models classify an input among listed options and large language models generate open-ended text, Clairic directly returns typed values: ranges, numbers, dates, lists, and distributions. The company describes the three categories as complementary phases, with extraction fueling the decisions and decisions fueling the final response generated.

Documented examples and reported tests

In the Cloudflare support email example, the model calculated a total of $701.10 from a stated price of $649.00 plus $52.10 in tax in the same step, calculated 11 days between the delivery of September 28, 2026 and the message date of October 9, 2026, listed the broken screen and dented corner as the exact ranges of the input, and returned null for a tracking number that the email never contained. In an example receipt photograph, Clairic extracted the order number, date and time of purchase, quantity, unit price, subtotal, sales tax, and tip with confidence scores and returned null for a discount that the receipt did not show.

As an estimate, the ad shows Clairic prices a home based on comparable sales text: The model returned a value of $418,000 with a median of $415,000 and an 80% range of $370,000 to $470,000 within the limits set by the caller. As for predictions, Cloudflare reports a 21-day test in which Clairic predicted the next day’s TSA airport checkpoint screening count from public series. The company says forecasts were 3.3% off on average, and that 17 of 21 actual counts fell within the model’s 80% range, with three of the four misses on Columbus Day weekend. For October 14, 2026, Cloudflare reported an actual count of 2,419,474 versus a predicted median of 2,419,283.

Cloudflare also reports a pipeline comparison pairing Clairic with Clef and the GLM-5.3 language model against a fully GLM-5.3 version of the same product damage claims workflow. According to the company, the waterfall completed in 19.3 seconds versus 61.7 seconds, 69% less time or about 3.2 times faster, and costs $6,032 per million requests versus $11,740, or 49% less. Clairic and Clef handled the extraction and decision in less than a second combined, where GLM-5.3 took 38 seconds for the same two steps, and Cloudflare says Gemma 4 and Qwen failed to derive values ​​such as repair cost and failed the task in the all-LLM version. Cloudflare has integrated the workflow in a live demo.

Preview of limitations and open weight plans

Cloudflare calls the current endpoint an early checkpoint and lists several limitations. One- and two-pass arithmetic such as totals, differences, conversions, and date arithmetic seems reliable, he said, but longer chains of dependent steps and aggregation over many table rows are difficult for a single-pass model without a notepad; recommends users rely on Clairic for prediction and handling arithmetic in code. Estimates need testing in the input, calibration is measured rather than guaranteed and ranges can be narrow on unfamiliar data, and forecasting is foundational rather than a substitute for dedicated time series models.

The company recommends using confidence scores, p_nulls, and ranges to decide what to automate and what to submit for further review, as the model’s responses are guaranteed to be well-formed and not incorrect. Cloudflare said it will continue to work on the arithmetic, estimation and forecasting aspects of the model with each checkpoint and that it plans to release Clairic’s open weights later. The name, he added, plays on paperwork, clarity and clairvoyance, and also serves as shorthand for Cloudflare’s AI regression model.