Cloudflare announced Clairic on October 19, a preview model that, instead of writing text, reads an input and returns typed values: numbers, dates, lists and probability distributions. The company calls it the first of a new class of “System One” extraction models. It runs on Cloudflare’s Workers AI platform, costs $0.10 per million input tokens, and is billed on input only because it generates no output tokens. Cloudflare says open weights will follow, without giving a date.

What Clairic does

A typical language model is asked to produce an answer token by token. Clairic works differently. A request pairs a “state”, meaning the text to read plus up to four optional images, with a map of typed fields. The supported field types are text, number, date, time, datetime, duration, list, estimate and forecast, each with optional plain-language instructions. The model fills every field in a single parallel forward pass, with no reasoning step, and Cloudflare says it typically responds in under a second. Fields that cannot fit beside the input in one 8,192-token pass are split across parallel passes that each see the whole input and are then merged.

Cloudflare’s own support-email example shows the idea: from a stated $649.00 price plus $52.10 tax, the model computed a $701.10 total and derived 11 days between a September 28, 2026 delivery and the claim date in the same pass, according to the report. This is closer to filling in a form than to chatting.

Built-in uncertainty

Every scalar answer carries p_null, the probability that the value is absent from the input, and a confidence score. 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. If an input does not determine a field, the model returns null and shows a “guess” it would otherwise have picked, rather than inventing a value. Cloudflare’s documentation advises using these signals to decide what to automate and what to send for review, and adds that answers are guaranteed to be well formed, not guaranteed to be right.

Availability and pricing

Clairic is in preview as @cf/cloudflare/clairic, callable from a Cloudflare Worker with env.AI.run() or via the REST API. Documentation lists a 65,536-token context window and vision support, with embedded images limited to four PNG, JPEG or WebP files of up to 4 MiB each. At $0.10 per million input tokens and no output charge, the cost of a call depends entirely on input size.

Where it fits among other models

Cloudflare positions Clairic beside its decision models. Clef and Clef-flash were released on October 1, 2026 and open-sourced on Hugging Face under Apache 2.0, post-trained on Qwen base models and compatible with the Jev API from TypeSafe. The company describes three complementary stages: extraction models pull values from messy input, decision models choose among listed options, and language models generate the final reply. In a damaged-product claims workflow, Cloudflare reports that a cascade of Clairic, Clef and the GLM-5.3 language model finished in 19.3 seconds against 61.7 seconds for an all-GLM-5.3 version, and cost $6,032 per million claims against $11,740. It also reported that Gemma 4 and Qwen failed to derive values such as repair cost in the all-language-model version. These are Cloudflare’s own results on its own workflow and have not been independently reproduced.

A second test forecast next-day TSA checkpoint screening counts from the public series: the company says forecasts were off by 3.3% on average and 17 of 21 actual counts fell within the model’s 80% interval. For October 14, 2026, the actual count was 2,419,474 against a predicted median of 2,419,283.

Limits Cloudflare acknowledges

The company calls the endpoint an early checkpoint. One- and two-step arithmetic such as totals, differences, conversions and date arithmetic seem reliable, but longer chains of dependent steps and aggregation across many table rows are hard for a one-pass model without a scratchpad, so it recommends doing arithmetic in code and using Clairic for the prediction. Estimates need evidence in the input, calibration is measured rather than guaranteed, and forecasting is basic, not a replacement for dedicated time-series models. Teams already using JSON modes should compare against structured outputs and function calling before moving workloads.

What to watch next

  • A date and license for the open-weights release.
  • Independent accuracy tests on invoices, forms and support tickets.
  • Whether other platforms copy the typed-field, no-generation design.
  • Improvements to arithmetic chains and forecasting in later checkpoints.
  • Pricing once the preview ends.

Key takeaways

  • Clairic fills typed fields in a single pass instead of generating text.
  • Preview price is $0.10 per million input tokens, with no output charge.
  • Each answer includes a null probability and confidence; estimates come with intervals.
  • Cloudflare’s speed and cost figures come from its own demo workflow.
  • Open weights are planned but undated.

Frequently asked questions

What is a “System One” model?

It is Cloudflare’s label for fast models that give an answer in one pass without a reasoning step, as opposed to language models that generate open-ended text.

Can Clairic replace a language model?

Not entirely. Cloudflare describes it as a complement: extraction first, decisions next, generated replies last.

Is it reliable for arithmetic?

Simple one- and two-step calculations seem reliable, according to Cloudflare, but long chains should be handled in code.