Cloudflare added Clairic to its Workers AI catalog on October 19. Workers AI is the service that lets developers call models from code without managing their own GPUs. The model identifier is @cf/cloudflare/clairic. A developer sends the text of an email, an invoice or a support ticket and may include up to four images. The developer also supplies a list of fields, each with a type and optional plain-language instructions.
The supported field types are:
- text, returned as a span with character offsets
- number, with optional decimal places and unit
- date, time or datetime in ISO form
- duration in a chosen unit
- list of matching text, number or date values
- estimate, a distribution inside declared bounds with a median and 80 percent interval
- forecast, one estimate per future step of a series
Clairic returns one typed answer per field, a confidence score, the character offsets of any extracted text, and a probability that the value is missing from the input. It can also compute. In an example Cloudflare published, an email stated a price of $649.00 and tax of $52.10. Clairic returned the sum of 701.10. It calculated the days between a delivery date and the date the customer wrote the message. When a requested field was absent the model returned null.
The model returns values in a single pass
Large language models generate tokens one by one. Clairic processes the input in a single forward pass and returns values of the types requested. Cloudflare charges only for the input at $0.10 per million input tokens. The context window is 65,536 tokens. Responses typically arrive in under a second.
Cloudflare describes Clairic as a System One model. The term comes from psychologist Daniel Kahneman’s book Thinking, Fast and Slow. TypeSafe AI introduced the label on September 15, 2026, when it released Jev, a model that selects among options a user lists and returns probabilities. Cloudflare launched its own decision model, Clef, on October 1. Clairic is the first model in the family that extracts or computes values.
Cloudflare reported a pipeline test
Cloudflare constructed a three-step example for a damaged-product support claim. Clairic extracted the facts. Clef selected the next action. An LLM wrote the short reply. The same LLM wrote the reply in both versions of the test. Cloudflare reported that the cascaded version finished in 19.3 seconds and the all-LLM version finished in 61.7 seconds. Cloudflare said the cascade cost 49 percent less per million claims. In the runs Cloudflare reported, both versions produced the correct facts and decision, while the LLM’s repair-cost estimate varied across runs and Clairic’s remained the same.
Those figures are Cloudflare’s reported results on its own example. They do not establish production accuracy on a buyer’s own documents. Extraction quality on the specific invoices, emails and tickets a team handles remains an evaluation question.
The model remains in preview. Cloudflare said it plans to release open weights. The company noted that multi-step arithmetic and aggregation across many rows of a table are not yet reliable, so those calculations should remain in code.
Ask the internal team which structured extraction jobs still go to a generative model. Then run Clairic against a sample of the actual documents those jobs handle and compare the confidence scores and null handling with the current process. The open weights Cloudflare has planned will matter only if the hosted preview does not meet the accuracy bar on those documents.
Cloudflare. “Introducing Clairic, a new System One class of extraction models.” Cloudflare Blog, 19 Oct. 2026, https://blog.cloudflare.com/clairic-extraction-model/.
Cloudflare. “clairic.” Cloudflare Workers AI docs, accessed 20 Oct. 2026, https://developers.cloudflare.com/workers-ai/models/clairic/.
TypeSafe AI. “Introducing System One Models & Jev.” TypeSafe AI Blog, 15 Sep. 2026, https://typesafe.ai/blog/introducing-system-one-models-and-jev.
