Futureweb

Distributed Cloud

Cloudflare's Clairic preview returns a typed value, not a paragraph. The weights are not open yet.

Cloudflare's Clairic preview returns a typed value, not a paragraph. Open weights are a stated intention, not a file you can fetch today. MarkTechPost retold it.

The short version

Cloudflare's changelog says Clairic, model id @cf/cloudflare/clairic, reads text or a few images and returns a typed value for each field you name, or null when the input does not decide it. The docs list it as a preview at $0.10 per million input tokens. Open weights are described as a later intention, not as a file you can fetch today.

A preview that answers with a value

Cloudflare's Workers AI changelog, dated October 19, 2026, says @cf/cloudflare/clairic is available as a preview. It calls Clairic an extraction model: you send a state and a schema of typed fields, and it returns one typed value per field. The blog the same period introduces that class of model. MarkTechPost's October 20 item is the retelling on the artificial-intelligence wire. The call still lands on Workers AI. Asked directly, Is MarkTechPost a second implementation? No. The October 20 item retells the launch. It is not a rerun, and it is not the open-weight release the changelog says is intended.when one region is the control plane.

Types the note is willing to name

Text comes back as a span of the input, with character offsets. Numbers honor an integer or decimal constraint. Dates, times, and durations come back in a structured form. Lists come back in input order. Estimates sit on a grid you bound with a minimum, a maximum, and a step. Forecasts are one estimate per future step, up to 64. The changelog says fields that do not fit beside the input in one pass are split across passes, so it claims no upper count of fields. Images, up to four, can sit with the text.the distributed-cloud desk.

Price, context, and the weight file

The model card lists a 65,536 token context and a unit price of $0.10 per million input tokens. The changelog says the model does not generate tokens in the usual sense, which is why an output-token price is the wrong picture. It also says Cloudflare intends to release the weights. Intends is not a download. A reader who wants a checkpoint they can fetch is not holding one because this preview exists. The document named in the sources is the record this page will keep.weights you can fetch.

What the preview says not to trust yet

The changelog separates a valid value from a correct one. It says short arithmetic is reliable and long chains are not, and it tells the reader to extract inputs and do the long math in code. It says forecasts follow a recent pattern and then flatten, and that they do not stand in for a time-series model trained on your own history. Intervals can run narrow on unfamiliar data. Those limits are the vendor's. They are also the reason a confidence score is part of the answer.an evaluation a second lab can run.

Null is an answer

The changelog says a field comes back null when the input does not determine it, and that a probability of null rides along. A missing purchase-order number is the example the docs use. Null is not a crash in that description. It is the model refusing to invent a span. Asked directly, What does the docs page charge? The model card says $0.10 per million input tokens, and the changelog's pricing note is an input charge rather than an output-token charge. Those are Cloudflare's prices.

What this page will not paste

It will not copy the request sample or a worked invoice total. The sample's numbers are Cloudflare's illustration. A reader's own document is not that invoice. The dependency is the type system and the preview label, not a demo total. Asked directly, What kinds of fields does the changelog list? Text as a span of the input, numbers, dates, times, durations, lists, estimates inside a min and max you set, and forecasts of up to 64 steps. Each scalar answer includes a probability that the input does not determine the value, and a confidence score.

What would change the preview label

An open-weight release with a license and a file. A price change on the model card. A second lab's error rate on a public set. A note that the endpoint left preview. MarkTechPost repeating the blog does not do any of those. October 19 is the changelog. October 20 is the retelling. Asked directly, What does the preview warn it does badly? The changelog says one- and two-step arithmetic is the reliable band, that long chains and big table sums are not, that forecasts several steps out flatten, and that a well-formed value is not the same as a correct one.Futureweb.

Workers AI is the control plane for this call

The model id lives in Cloudflare's catalog. A request that uses it is a request to that network, with that account's token. Typed output does not move the weights onto the caller's machine. On-device inference is a different brief. Asked directly, Where does the preview run? On Workers AI, under the model name @cf/cloudflare/clairic. The changelog is dated October 19, 2026. Calling it is a request to Cloudflare's network, not a weight file on a desk.what on-device inference keeps.

What stays on the record

October 19, a preview id, typed fields, null when the input does not decide, a published input price, a stated intention to open the weights later, and an October 20 retelling. The short answer this page will stand on is this: Cloudflare's changelog says Clairic, model id @cf/cloudflare/clairic, reads text or a few images and returns a typed value for each field you name, or null when the input does not decide it. The docs list it as a preview at $0.10 per million input tokens. Open weights are described as a later intention, not as a file you can fetch today.

Sources

The reports this brief is filing. Futureweb did not republish them.

  1. Cloudflare changelog
  2. Cloudflare blog
  3. MarkTechPost

Questions

Where does the preview run?

On Workers AI, under the model name @cf/cloudflare/clairic. The changelog is dated October 19, 2026. Calling it is a request to Cloudflare's network, not a weight file on a desk.

What kinds of fields does the changelog list?

Text as a span of the input, numbers, dates, times, durations, lists, estimates inside a min and max you set, and forecasts of up to 64 steps. Each scalar answer includes a probability that the input does not determine the value, and a confidence score.

What does the docs page charge?

The model card says $0.10 per million input tokens, and the changelog's pricing note is an input charge rather than an output-token charge. Those are Cloudflare's prices.

What does the preview warn it does badly?

The changelog says one- and two-step arithmetic is the reliable band, that long chains and big table sums are not, that forecasts several steps out flatten, and that a well-formed value is not the same as a correct one.

Is MarkTechPost a second implementation?

No. The October 20 item retells the launch. It is not a rerun, and it is not the open-weight release the changelog says is intended.