Cloudflare announced a preview of Clairic on October 20, a model hosted on Workers AI that returns typed values for named fields in one pass instead of generating text. The company says open weights are coming soon.
What happened
According to Cloudflare’s announcement, a request is a block of text, plus up to four optional images, and a map of typed fields. Each answer comes back with a confidence score and the probability that the value is absent, and a field the input does not contain returns null. It can also compute, estimate and forecast values. Billing covers input tokens only. It follows the decision models Jev and Clef in what Cloudflare calls the System One class, the idea behind TypeSafe’s Jev.
Why it matters
Cloudflare ran a damaged-product claim through two pipelines. Clairic, Clef and GLM-5.3 finished in 19.3 seconds at $6,032 per million claims. GLM-5.3 handling all three steps took 61.7 seconds at $11,740. By our arithmetic that is about $0.006 against $0.012 per claim. Both pipelines got the facts and the decision right, but Clairic’s repair estimate was identical across runs while GLM-5.3’s moved between $250 and $220.
On a public series, Cloudflare asked Clairic for the next day’s TSA checkpoint count for 21 days. Forecasts were off by 3.3% on average and 17 of the 21 actuals fell inside the 80% interval. Three of the four misses fell on the Columbus Day weekend. The Decoder adds that Cloudflare says Clairic cannot yet handle longer chains of calculations reliably. Output tokens are a growing share of model bills, a point that runs through our look at Sonnet 5.5 and token efficiency.
What to ask
Both numbers come from one vendor-built pipeline on one task. A team with high-volume invoice or claims extraction should rerun it on its own documents and record the null rate and how many values fall below its confidence threshold before moving any workload off an LLM.
Source: Cloudflare Blog
