Cloudflare launches model that extracts numbers and text without writing answers

Cloudflare announced the Clairic model on October 19, 2026. It is the first in a new class that the company calls System One extraction models. It is available as a preview on Cloudflare's Workers AI platform.

The model is built to pull specific information out of a text, such as an email or an insurance claim. It can also calculate, estimate and forecast values. The user defines in advance which fields should be filled in, and gets answers back with a confidence score. In other words, it tells you how certain the model is.

No text, just answers

The difference between Clairic and the familiar chat models is that it processes input in a single forward pass, without generating tokens. Tokens are the small pieces of text a language model uses to write answers. When the model doesn't write out the answer word by word, it is faster. According to Unite.AI, response time is typically under one second.

Cloudflare's own numbers

Cloudflare claims that Clairic delivers higher accuracy, lower latency and lower cost than traditional language models. In a test of a claims processing workflow, it took 19.3 seconds from start to finish, compared with 61.7 seconds for a solution based solely on language models. That is 69 percent faster. The cost was $6,032 per million claims versus $11,740, or 49 percent less.

In a demo forecasting the number of passengers at US TSA checkpoints over 21 days, the average error was 3.3 percent. 17 of 21 actual values fell within the model's 80 percent interval.

All figures come from Cloudflare's own tests and demos and have not been independently verified.

Pricing and limitations

According to Unite.AI, it costs $0.10 per million input tokens, and there is no charge for output because the model does not generate tokens. The context window, meaning how much the model can take in at once, is 65,536 tokens, and it supports up to four images. This information comes only from Unite.AI and should be checked against Cloudflare's documentation.

Cloudflare itself says that longer chains of reasoning and calculations across many rows are unreliable without external computation. Estimates require grounding in the input, and calibration is measured but not guaranteed on unfamiliar data. According to Cloudflare, forecasting is basic and no substitute for specialized time series models. The answers are guaranteed to have the right format, but not necessarily the right content.

Cloudflare plans to release an open weights version, which would let others download the model itself. No date has been given.