Cloudflare published a preview of Clairic on October 19, 2026, the first of a class it calls System One extraction models. It runs on Workers AI under the name @cf/cloudflare/clairic, and it returns typed values read from text and images instead of writing a reply.
What was actually released
Clairic is not a chat model with a JSON mode bolted on. A request carries a state, the text to read plus up to four optional images, and a map of named fields. Each field declares a type and, if needed, a plain-language instruction. The model answers in one pass and never generates a token.
| Field type | What comes back |
|---|---|
| text | a span of the input, with offsets |
| number | read or computed, decimals honored |
| date, time, datetime | valid dates and times, in ISO form |
| duration | a length of time, in your unit |
| list | text, number or date, in input order |
| estimate | a distribution inside your bounds, linear or log |
| forecast | one estimate per future step |
Because there are no output tokens, every answer is a value of the declared type or null. A field the input lacks comes back as null, with the probability that it is null, rather than as a plausible guess. Every scalar answer carries a confidence score, and every estimate a median with a p10 to p90 interval.
The field count is unbounded: fields that do not fit beside the input are split into parallel passes that each see the whole input. Thirty fields costs about the same as three.
What ships on October 20
- A preview endpoint on Workers AI, which Cloudflare calls an early checkpoint.
- Text and up to four images per request.
- Seven field types, covering extracted, computed, estimated and forecast values.
- Open weights promised, with no date attached.
What the claim rests on
The post contains one measured comparison, and it measures a pipeline rather than the model. Cloudflare ran a damaged-product claim through two versions: one where Clairic extracts the facts, Clef makes the decision and GLM-5.3 writes the reply, and one where GLM-5.3 performs all three steps from the same email, policy and prompts.
| Measure | Clairic, then Clef, then GLM-5.3 | GLM-5.3 for every step |
|---|---|---|
| Latency, end to end | 19.3 s | 61.7 s |
| Extraction | Clairic, 0.46 s | GLM-5.3, 16.1 s |
| Decision | Clef, 0.22 s | GLM-5.3, 21.9 s |
| Reply | GLM-5.3, 18.6 s | GLM-5.3, 16.4 s |
| Input tokens | 701, at $184 per million claims | 390, at $546 per million claims |
| Output tokens | 1,329, at $5,848 per million claims | 2,544, at $11,194 per million claims |
| Cost per million claims | $6,032 | $11,740 |
Two readings matter more than the headline. First, the all-LLM run puts almost all its cost in output tokens: $11,194 of $11,740, which is 95% of the bill by our arithmetic. Clairic’s contribution is not a low price per call but the removal of a generation step that used to run over the raw email. Second, the input line moves the other way: the cascade sends more input tokens, 701 against 390, at a lower input cost, $184 against $546. The post publishes no per-call rate for Clairic or Clef, so that line cannot be rebuilt from token counts. Read the 49% as Cloudflare’s total, not as a rate you can drop into a calculator.
The rest of the evidence is thinner. Cloudflare reports that Gemma 4 and Qwen failed at the same task in the all-LLM version, on its own harness, with no document set published. The forecast test is more useful because it runs on a public series: 21 consecutive mornings of TSA checkpoint counts, a mean error of 3.3%, and 17 of 21 counts inside the published 80% interval. Three of the four misses fall on the Columbus Day weekend, which is the honest part. The model reads the shape of a series and knows nothing about holidays.
The architecture makes every answer well formed, not necessarily correct.
What Cloudflare does not publish
- No price per call or per token for Clairic on the announcement page.
- No parameter count, architecture detail, training data description or evaluation set.
- No date for the open weights, and no license.
- No accuracy table on a public corpus of documents, and no run by anyone outside Cloudflare.
- No throughput figure or rate limit for the preview endpoint.
On this page
What it changes for teams extracting data at volume
Extraction is the easiest part of a document pipeline to overpay for. Invoices, receipts, claims and support emails all end in the same shape: a number, a date, a name, a category. Sending that shape to a frontier model means paying for reasoning nobody asked for, on every field of every document, then paying again in checks because a generated digit can be wrong.
Three things change when a typed model sits at that step. The bill moves from output to input, five times cheaper per token at Anthropic’s cheapest small model, $0.10 against $0.50 per million (Claude Haiku 5.5). Confidence becomes a routing decision: the probability of null and the interval tell you which items deserve a second look, and one example shows a missing tracking number returning null at 0.42 instead of a convincing fake. And arithmetic moves into code, because Cloudflare says one-pass models struggle beyond two dependent steps.
The preview has its own cost: one endpoint on one provider, no published rate, and weights still a promise. A team that rewrites its pipeline around this checkpoint bets on a model its vendor calls early, measured on one email.
Output tokens are 95% of the all-LLM bill, and that is the line a typed model removes.
The Neural Minute take
The useful news is not a model. It is the split. Extract with a typed model, decide with a typed model, write with a generative one, and let deterministic code do the arithmetic between them. It is worth adopting this week and does not require Clairic: every frontier provider returns a typed field with a confidence score when asked, and routing low-confidence items to a second look is what removes the manual checking. Most pipelines do the opposite, paying one model to read, reason and write, then a human to verify.
What Clairic adds is a price signal in the right direction. A model that bills only input, answers in 0.46 seconds and returns null instead of inventing a missing field is the correct shape for the boring half of document work.
Our decision for now is not to migrate. The endpoint is a preview, the rate card is unpublished, and the caveats in the post, weak multi-step arithmetic, basic forecasting, calibration that can run narrow, remove precisely the tasks most teams want to hand over first. The open weights are the event to watch. Weights would let a pipeline run extraction inside its own infrastructure, at a cost unrelated to a per-call rate, and that is when 49% becomes a figure you can test on your own documents rather than Cloudflare’s.
17 of 21 daily forecasts landed inside the published 80% interval on one public series.
Until then, the measurable move is free. Count what share of your model bill is generation you did not need. If it looks like the 95% above, the model you picked was never the problem.
Sources
- Introducing Clairic, a new System One class of extraction models, Cloudflare, published October 19, 2026. The figures, the quote and the limitations in this article are read there.
- Anthropic prices Haiku 5.5 at $0.10 per million tokens, Neural Minute, for the input and output rates in the comparison above.