Skip to content

CLOUDFLARE 2026-10-19 Tier 1

Read original ↗

Introducing Clairic, a new System One class of extraction models

Summary

Cloudflare announced a preview of Clairic (@cf/cloudflare/clairic on Workers AI), the first model in a new class of System One extraction models — a sibling to the decision models Clef / Clef-flash and to TypeSafe's Jev. Where a decision model picks among a fixed set of options with probabilities, an extraction model answers with typed values that are extracted from the input, computed from it, or predicted beyond it. Like the decision models, Clairic is System One — fast, typed, cheap, answering in a single parallel forward pass with no generated output tokens, so it is billed only for input and typically responds in under a second where a reasoning LLM takes several. Because it generates no tokens, every answer is a value of the declared type (or null), inside the bounds the caller declares — it structurally cannot emit a mistyped or hallucinated answer, and ships confidence / p_null scores and calibrated intervals so callers can decide what to automate vs route for review. The post positions extraction models as the first stage of a Clairic → Clef → LLM cascade ("extract values → decide → act") and benchmarks that cascade at 3.2× faster and 49% cheaper than an all-LLM pipeline on a damage-claim workflow.

Key takeaways

  • A new model class: System One extraction models. Until now every System One model was a decision model (Jev, Clef) taking an input state + a list of questions and returning probabilities over listed options. Clairic answers with values directly extracted from, computed from, or predicted beyond the input — the regression/extraction counterpart to the classification framing of decision models (Source: sources/2026-10-19-cloudflare-introducing-clairic-extraction-model).
  • Single parallel forward pass, no output tokens. Like a decision model, Clairic "only processes input tokens with a single forward pass, and thus has no charge for output tokens" — faster, cheaper, and (for its tasks) more accurate than an LLM. There is "no reasoning and no output tokens — everything computes in a single forward pass" (concepts/hot-path, concepts/latency-critical-vs-latency-tolerant-workload).
  • Four kinds of answer. Extracted values (text copied from the input, returned as a character-offset span + confidence), computed values (numbers, dates, times, datetimes, durations, lists — read or worked out from the input, honoring type rules like two-decimal cents), estimated values (a distribution inside caller-declared [min, max] bounds on a step grid, with a median + 80% interval), and forecasted values (one estimate per future step of a series, each with its own 80% interval).
  • Typed-or-null, never a plausible-looking guess. Each scalar answer carries p_null (probability the value isn't in the input) and confidence; a field not present comes back value: null (with a guess showing what the model would have picked). In the worked example a missing tracking_number returns null with p_null: 1 rather than fabricating one — the structural answer to concepts/llm-hallucination and the strongest form of concepts/structured-output-reliability ("it can't produce a mistyped answer or 'hallucinate' in the same way as an LLM").
  • Unlimited fields via server-side parallel passes. A request is a state (text + up to four optional images) and a map of typed fields; "if they don't fit beside the input in one pass, the server splits them into parallel passes that each see the whole input, then merges the answers," so asking for thirty fields "costs about the same as asking for three."
  • Images too. Clairic extracts typed values from a photographed receipt (order number, datetime, quantity, unit price, subtotal, tax, tip) with per-field confidence, correctly returning null for a non-present discount (patterns/multimodal-content-understanding).
  • Estimates and forecasts carry calibrated intervals. On comparable home sales Clairic returns sale_price as a bounded distribution (value 418000, median 415000, p10 370000, p90 470000); on a 21-day TSA traveler-count forecast it was off by 3.3% on average with 17/21 actual counts inside the 80% interval (right at the expected calibration). The misses were holiday-weekend days the series didn't signal — Clairic "forecasts from the patterns in the series but doesn't account for events the series doesn't show." (interval width = data confidence.)
  • The cascade: extract → decide → act. Extraction turns raw input into typed values, a decision model (Clef) turns values into a decision, an LLM turns the decision into words. Each step gets cleaner input than the last; the LLM never sees the raw email (only the typed decision + facts), cutting concepts/prompt-injection surface and the number of values it can misuse. Canonical cheap specialist stages gating an expensive generalist + typed decomposition instance.
  • Measured cascade win. Clairic → Clef → GLM-5.3 vs all-GLM-5.3 on a damage-claim workflow: 19.3 s end-to-end vs 61.7 s (3.2× / −69%) and $6,032 vs $11,740 per million claims (−49%); Clairic extract 0.46 s + Clef decide 0.22 s vs GLM-5.3's 16.1 s + 21.9 s for the same two steps. Clairic's repair estimate was identical on every run; GLM-5.3's varied ($250/$250/$220). Gemma 4 and Qwen failed the all-LLM version — "unable to deduce values like repair_cost from the inputs," underscoring LLMs' weakness at data extraction.
  • Preview caveats (Cloudflare's own). Reading is strong but multi-step arithmetic isn't yet ("have Clairic predict the value and let code do the math"); estimates need in-input evidence; calibration is measured, not guaranteed (intervals can run narrow on unfamiliar data); forecasting is basic (not a replacement for a dedicated time-series model); and callers should use confidence / p_null / intervals to decide what to automate. Hosted preview now; open weights "soon."

Systems / concepts / patterns extracted

  • New system page: Clairic — the extraction model (primary subject; substantively described — request model, four answer types, parallel passes, serving on Workers AI edge GPUs).
  • Enriched systems: Clef (its decision-model sibling + the shared System One lineage + the extract→decide→act cascade), Workers AI (hosts Clairic).
  • Named in passing (systems frontmatter, no page): Jev / TypeSafe (the originating System One decision model Clairic is API-shaped against), GLM-5.3 (@cf/zai-org/glm-5.3, the cascade's writer LLM), Gemma 4, Qwen (the all-LLM baselines that failed the extraction task).
  • Concepts (mapped into existing): concepts/structured-output-reliability (typed-or-null output a consumer can always parse), concepts/llm-hallucination (can't hallucinate by construction — null instead of a guess), concepts/hot-path + concepts/latency-critical-vs-latency-tolerant-workload (single-pass sub-second inference in the agent path), concepts/human-in-the-loop (confidence/p_null route low-trust answers for review), concepts/knowledge-distillation (a specialized non-generative model standing in for a general LLM), concepts/prompt-injection (the LLM never sees raw input in the cascade).
  • Patterns (mapped into existing): patterns/cheap-approximator-with-expensive-fallback (cheap typed specialist gating the expensive LLM), patterns/specialized-agent-decomposition (extract/decide/act typed-stage decomposition), patterns/multimodal-content-understanding (typed extraction from images).
  • Prose-only (single-source named ideas; Lint candidates at ≥3 citations): System One extraction model, typed-value-or-null-never-a-guess, character-offset span extraction, estimate-as-bounded-distribution, server-side parallel-field-pass merge, extract-decide-act cascade, regression-as-System-One — Clairic-specific framings, recorded here and on the Clairic page, not minted as concept/pattern pages.

Operational numbers

  • Cascade end-to-end: 19.3 s (Clairic→Clef→GLM-5.3) vs 61.7 s (all GLM-5.3) — 3.2× / −69%.
  • Per-step: Clairic extract 0.46 s; Clef decide 0.22 s; GLM-5.3 reply 18.6 s (vs GLM-5.3 extract 16.1 s + decide 21.9 s + reply 16.4 s).
  • Cost per million claims: $6,032 (cascade) vs $11,740 (all-LLM) — −49%. Input tokens 701 ($184/M) vs 390 ($546/M); output 1,329 ($5,848/M) vs 2,544 ($11,194/M).
  • Forecast: 21-day TSA traveler counts off 3.3% on average; 17/21 inside the 80% interval; one day predicted median 2,419,283 vs actual 2,419,474.
  • Request shape: state text + up to 4 images; unlimited fields (split into parallel passes); scalar answers carry p_null + confidence; estimates carry median + p10/p90 (all 99 percentiles on request).

Caveats

  • Preview / early checkpoint. Cloudflare expects it to improve; open weights are a future release. Benchmarks are Cloudflare-reported (single damage-claim workflow + one public TSA series); no independent reproduction.
  • Known weak spots named by Cloudflare: multi-step/dependent arithmetic and many-row table aggregation are unreliable in one pass (do the math in code); estimates need in-input evidence; calibration is measured-not-guaranteed and can run narrow on unfamiliar data; forecasting is basic (non-critical pattern matching only, no long-horizon/holiday awareness).
  • Trust the scores, not the architecture. The single-pass design makes every answer well-formed, not necessarily correct — confidence / p_null / interval width are the intended gate for automation vs human review.

Source

Last updated · 854 distilled / 2,448 read