Signals
Cloudflare Lets Agents Skip the LLM for Data Extraction
Agents no longer need a frontier LLM to pull typed fields out of messy input. Cloudflare's Clairic, a preview on Workers AI, returns each field with confidence and p_null in one pass, bills no output tokens, and extracted in 0.46s vs 16.1s for GLM-5.3 in Cloudflare's test. Keep long math chains in code.
What Changed
- A new model class, live today. Cloudflare released Clairic as a preview on Workers AI as @cf/cloudflare/clairic, the first "System One extraction model", and says it intends to release open weights soon.
- Typed fields instead of generated JSON. You send text plus up to four images and a map of fields typed text, number, date, time, datetime, duration, list, estimate or forecast. Every scalar carries confidence and p_null, and a value not in the input comes back null.
- Billed on input only. Clairic answers in one parallel forward pass and generates no tokens, so there is no output charge. The model page lists $0.10 per million input tokens, and thirty fields cost about the same as three.
- Computed and estimated values, not just copied ones. Clairic added price and tax into a total and worked out days since delivery. Estimate fields return a median and an 80% interval inside bounds you declare.
- A sourced end-to-end benchmark. A damage-claim pipeline of Clairic, Clef, then GLM-5.3 took 19.3 s against 61.7 s for GLM-5.3 on every step, about 3.2x faster, and cost $6,032 against $11,740 per million claims, 49% less.
Who Should Care
- Agent builders on Workers whose agents read inbound email, tickets, receipts or PDFs and hand an LLM a "return these fields as JSON" prompt with a retry loop behind it.
- Teams running high-volume structured-output calls (invoice totals, delivery dates, order IDs, SLA durations) where output tokens and multi-second latency are the bulk of the bill.
- Anyone feeding untrusted text to an agent that takes actions. Cloudflare's design has decision and reply steps work from typed values, so the LLM never sees the raw email, narrowing the prompt injection surface.
Who Should Not Care
- Teams whose LLM calls are genuinely open-ended: writing, summarizing, explaining, or fixing code. Clairic does not generate text.
- Workloads that are mostly long arithmetic chains or aggregation across large tables, which Cloudflare says this checkpoint does not handle well yet.
- Teams with a tuned, dedicated time-series model in production. Cloudflare says Clairic's forecasting is basic and does not replace one.
The Verdict
Move your highest-volume LLM extraction call to Clairic on a branch this week. Pick the one call that turns raw input into fixed fields and map each field to a Clairic type. Route answers by confidence and p_null: auto-accept above your threshold, send the rest to review or your existing LLM path. Run both paths side by side on the same real inputs and compare latency, cost per thousand and field accuracy before you switch traffic. Keep any multi-step math or cross-row totals in code, because Cloudflare says this preview reads well but does not yet chain calculations reliably.
Take This To Your Agent
Copy a ready-to-paste investigation handoff that asks your coding agent to check this update against your product and recommend what to do next.