Cloudflare on 19 October 2026 previewed Clairic, a new “System One” extraction model hosted on Workers AI as @cf/cloudflare/clairic. Instead of generating text token by token, Clairic answers in a single forward pass, returning typed values (names, dates, sums, estimates and forecasts) with confidence scores — and charging only for input. For developers who now send invoices, support emails and receipts to frontier LLMs just to get JSON back, the pitch is faster, cheaper and less error-prone extraction.
The release is a preview, not a finished product. Cloudflare says it intends to publish open weights “soon”, but for now Clairic is available only as a hosted endpoint, and every speed, cost and accuracy figure in the announcement comes from Cloudflare’s own tests.
In brief
- Cloudflare previewed Clairic, an extraction model on Workers AI that returns typed values in one pass with confidence and null probabilities; open weights are planned but not yet released.
- In Cloudflare’s damaged-product claim test, a Clairic → Clef → GLM-5.3 pipeline ran in 19.3 s versus 61.7 s for GLM-5.3 alone, at $6,032 versus $11,740 per million claims.
- Cloudflare calls the endpoint an “early checkpoint” and says multi-step arithmetic “isn’t yet” strong; its figures are not independently verified.
A model that fills fields instead of writing prose
Cloudflare credits the “System One” idea to TypeSafe, whose Jev decision model kicked off the category. Decision models such as Jev and Cloudflare’s own Clef pick among options a developer lists in advance. Clairic is the first of a second class: it returns values that cannot be listed beforehand, such as an invoice total or a repair-cost estimate.
A request contains a state (text plus up to four optional images) and a set of typed fields. Clairic returns text spans with character offsets, computed numbers, dates, durations and lists, estimates with a median and an 80% interval (p10 to p90), and forecasts for time series. Every scalar answer carries p_null and a confidence score, and a missing value comes back as null rather than a plausible-looking guess.
Because it generates no tokens, Cloudflare says, Clairic cannot produce a malformed answer. The company is careful to add that the architecture makes every answer “well formed, not necessarily correct”, so developers should use the confidence and interval data to decide what to automate.
Cloudflare’s numbers, Cloudflare’s test
The headline comparison is a three-step claims pipeline. Clairic extracts the facts from a customer email, Clef decides between refund and repair, and GLM-5.3 writes the reply. Cloudflare ran it against a version in which GLM-5.3 did all three steps.
| Step | Clairic → Clef → GLM-5.3 | GLM-5.3 only |
|---|---|---|
| Extract | 0.46 s | 16.1 s |
| Decide | 0.22 s | 21.9 s |
| End to end | 19.3 s | 61.7 s |
| Cost per million claims | $6,032 | $11,740 |
On those figures, Cloudflare says the cascade took 69% less time, about 3.2 times faster, at 49% lower cost. It also says Gemma 4 and Qwen failed the all-LLM version. In a separate forecasting test on daily TSA airport screening counts, Clairic was off by 3.3% on average over 21 days, and 17 of 21 actual counts fell inside its 80% interval. Most misses landed on the Columbus Day weekend, which the series alone did not signal.
The caveats Cloudflare puts first
Cloudflare describes the current endpoint as “an early checkpoint” and expects it to improve. Its own list of limits is blunt: reading is strong, but multi-step arithmetic “isn’t yet”. One- and two-step calculations look reliable, while longer chains and aggregation across many table rows are hard for a one-pass model, so Cloudflare suggests letting code do the maths.
It also says estimates need evidence in the input, that calibration is “measured, not guaranteed” and can run narrow on unfamiliar data, and that forecasting is basic and no replacement for dedicated time-series models. Those limits matter for anyone tempted to drop Clairic straight into finance or compliance workflows.
Our read
Clairic extends the System One idea from choosing to extracting, and the typed-output, confidence-first design makes sense for audit-heavy pipelines where every value must be traceable to its source span. The catch is that all the evidence so far is Cloudflare’s own demo and benchmarks, with the arithmetic limits Cloudflare itself flags. The test to watch is the open-weight release: once outside developers can run and benchmark Clairic themselves, the 3.2× claim will either hold up or shrink.
How we sourced this
- Primary: Cloudflare blog by Alex Reneau, Michelle Chen and Kevin Flansburg (19 October 2026). All latency, cost, forecasting and failure figures are Cloudflare’s own and have not been independently reproduced.
- Clef is described only as Cloudflare’s existing decision model; no launch date for it is given here.
Sources
Related stories
- TypeSafe AI, Maker of Jev, Raises $870 Million at a $7.5 Billion Valuation
- Vercel Says Jev Was the Fastest-Adopted Model on AI Gateway
- RAG vs Jev+RAG Explained: What the Judgment Gate Adds
- IBM Releases Granite Switch 4.1 With Decision Adapters and Fast Speech Recognition
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
