Clairic: A System One Extraction Model
What is the invoice total, when was this delivered, how many days late, what will this repair cost — the same structured calls keep hitting frontier LLMs, expensively and slowly. Clairic takes another route: one forward pass, no reasoning, no token generation, pulling values out of unstructured input, computing, estimating and even forecasting them with 80% intervals and confidence scores, billed on input only and typically under a second. It is in preview on Workers AI as @cf/cloudflare/clairic, with open weights promised soon
Tool Interface
Interactive tool will be available soon
Features
- ✓ The first member of the System One extraction class, sitting alongside decision models such as Jev and Clef rather than being an LLM
- ✓ Four output kinds: extracted values (text span plus character offsets plus confidence), computed values (numbers, dates, times, durations, lists following type rules such as two decimals for cents), estimated values (a distribution inside declared bounds with median and 80% interval) and forecasted values (one estimate per step ahead, each with an 80% interval)
- ✓ Answers in a single pass with no reasoning and no token generation — the post contrasts this with an LLM taking several seconds for the same step
- ✓ Billed on input only, so output costs nothing, which suits high-frequency structured calls
- ✓ Available in preview on Workers AI as @cf/cloudflare/clairic, with open weights promised soon
- ✓ Builds on the System One idea popularised by TypeSafe: fast, precise and typed, aimed exactly at what LLMs do slowly and expensively
How to Use
- Split tasks by output type first: a number or a typed value goes to Clairic, a pick from a fixed set goes to Clef, open-ended text still belongs to an LLM
- Call @cf/cloudflare/clairic in Workers AI, declaring the fields and types you want; estimated values also need explicit bounds and a grid
- Set thresholds on the returned confidence and 80% intervals, routing low-confidence records to human review or a fallback LLM
- Watch the preview closely and re-evaluate self-hosting cost and data boundaries once the open weights land
FAQ
What is Clairic?
A System One extraction model that Cloudflare is previewing on Workers AI under the identifier @cf/cloudflare/clairic. It extracts, computes, estimates and forecasts values from unstructured input, with the stated goal of being more accurate, faster and cheaper than traditional LLMs. Source: https://blog.cloudflare.com/clairic-extraction-model/
How does it differ from an LLM?
It answers in a single forward pass with no reasoning and no token generation, so responses typically land under a second — the post notes an LLM needs several seconds for the same step. Billing covers input only, not output.
What kinds of results can it return?
Four: extracted values (text spans with character offsets and confidence), computed values (numbers, dates, times, datetimes, durations, lists following type rules such as two decimals for cents), estimated values (a distribution on the bounds and grid you declare, with median and an 80% interval) and forecasted values (one estimate per step ahead, each with an 80% interval).
How does it relate to Jev and Clef?
Jev and Clef are decision models: you hand over an input state plus a list of questions and get probabilities back. Clairic is a new class within the same System One family — extraction models that return values read from, computed from or predicted beyond the input.
Can I use it today, and is it open source?
The preview is usable now, hosted on Workers AI, with docs at https://developers.cloudflare.com/workers-ai/models/clairic/ . Cloudflare says it intends to release open-weights versions soon.