On Oct. 19, Cloudflare put Clairic into preview on Workers AI as @cf/cloudflare/clairic. Cloudflare Clairic is an extraction model that reads unstructured text and returns typed values in a single pass, without generating tokens. Cloudflare says it intends to release open weights "soon" and gave no date.
It targets repetitive structured-output calls, such as an invoice total, a delivery date or a repair cost, that teams now send to frontier LLMs.
What Clairic does
Cloudflare calls it the first "System One" extraction model. System One models are fast and answer in one pass, unlike LLMs that reason step by step. Cloudflare says Clairic can extract, compute, estimate and forecast values from messy input, and typically responds in under a second.
Estimate and forecast answers come back with a median and an 80% interval. That is the range meant to hold the true value about 8 times out of 10, though Cloudflare says calibration is measured, not guaranteed. For estimates, Cloudflare says all 99 percentiles are available on request.
Plain number answers return a value, a confidence score and a p_null, the chance the value is not in the input. Clairic also pulls typed values, meaning answers returned as a set type such as a name, date, duration or list.
What Cloudflare's test shows
In Cloudflare's own test, one example pipeline handled a damaged-product support email. Clairic extracted the facts, Clef made the decision and GLM-5.3 wrote the reply. Clef is Cloudflare's companion decision model, which picks among listed options and returns a probability for each. The comparison ran GLM-5.3 on every step.
| Measure | Clairic, Clef, GLM-5.3 | GLM-5.3 only |
|---|---|---|
| End-to-end time | 19.3 s | 61.7 s |
| Extraction step | 0.46 s | 16.1 s |
| Cost per million claims | $6,032 | $11,740 |
That is 69% less time and 49% less cost, by Cloudflare's math. Cloudflare says Gemma 4 and Qwen failed the task in the all-LLM version. It also says Clairic's repair estimate was identical across runs, while GLM-5.3's varied ($250, $250, $220).
In a separate test on public TSA checkpoint data, Cloudflare asked for each next day's count over 21 days. Forecasts were off by 3.3% on average, and 17 of 21 actual counts landed inside the 80% interval. Most misses fell on Columbus Day weekend.
Price and how to try it
Cloudflare's pricing page lists Clairic at $0.100 per million input tokens as of 6:50 PM ET Oct. 19. The same page gives 9,091 neurons per million, Cloudflare's billing unit on Workers AI. The model charges only for input. Call it from a Worker with env.AI.run() or through the REST API. The model page has the docs.
Where it falls short
Cloudflare calls this an "early checkpoint" and lists its own caveats:
- One- and two-step calculations seem reliable. Longer chains of dependent steps, and aggregating across many rows, are not yet. Cloudflare says to have Clairic predict the value and let code do the math.
- Estimates need evidence in the input, such as comparables or history.
- Calibration is measured, not guaranteed, and intervals can run narrow on unfamiliar data.
- Forecasting is basic and is no replacement for models trained on a specific series.
Cloudflare says to use confidence, p_null and the intervals to decide what to automate and what to send for review.
Who should try it
Going by Cloudflare's stated limits, developers with high-volume extraction pipelines already on Workers AI have the clearest reason to test it against their own documents. Anyone who needs reliable long arithmetic or serious forecasting, or who wants open weights first, has reason to wait.
For more on AI developer infrastructure and cost, see Amazon's TorchNeuron. If you wire a model like this into an app, our guide on rotating AI API keys and setting spending limits covers keeping the bill in check.
