Cloudflare's Clairic Answers With Typed Values, Not Tokens

Cloudflare announced Clairic on October 19, and the framing is as interesting as the model: not a smaller language model, but a different category the company calls “System One” extraction models, available in preview on Workers AI as @cf/cloudflare/clairic. Where an LLM answers by generating tokens one at a time, Clairic runs a single forward pass, emits no output tokens at all, and returns typed values: extracted text spans with character offsets, computed numbers and dates, estimated distributions, even short-horizon forecasts, each with a confidence score and a probability-of-null. Thirty fields cost about the same as three, and responses typically land in under half a second.

What an extraction model actually returns

Every answer falls into one of four categories. Extracted values cite their location in the source. Computed values derive numbers, dates, durations, and lists from the input. Estimated values return distributions within declared bounds, with a median and an 80% interval (all 99 percentiles on request). Forecasted values extend that to per-step futures, each with its own interval. The demo set includes a blurry receipt photo yielding prices and quantities, with a correct null for a missing discount, and a 21-day TSA checkpoint-volume forecast that averaged 3.3% error with 17 of 21 actuals inside the 80% interval, missing exactly where a holiday weekend broke the pattern. Because there are no generated tokens, the hallucination profile changes shape: the model can still be wrong, but it cannot produce fluent nonsense, and every answer ships with calibration metadata you can gate on.

The benchmark: a claims pipeline, end to end

Cloudflare’s headline comparison runs an insurance claims workflow three ways. The all-GLM-5.3 pipeline took 61.7 seconds and cost $11,740 per million claims; a Clairic-then-Clef-then-GLM pipeline took 19.3 seconds and $6,032, a 3.2x latency win and 49% cost reduction, with Clairic extracting in 0.46 seconds and Clef, Cloudflare’s decision model, deciding in 0.22. The determinism detail is the one practitioners should weight most heavily: Clairic’s repair estimate was identical across runs where the LLM’s varied between $250 and $220, and Cloudflare notes outright that Gemma 4 and Qwen failed the all-LLM extraction task entirely. The architecture being proposed is a pipeline of specialized stages, extraction models turn raw input into typed values, decision models turn values into decisions, LLMs turn decisions into words, with each handoff cleaner, more auditable, and more injection-resistant than free-text passing between general models.

The honest limits, and what to watch

The post is unusually candid about boundaries: multi-step arithmetic and cross-row aggregation work reliably only at one or two steps, estimates require evidence actually present in the input, calibration is measured rather than guaranteed, and the forecasting is explicitly not a substitute for dedicated time-series models. Availability is preview on Workers AI via env.AI.run() or REST, with open weights promised soon, which matters because the extraction layer is exactly the tier where self-hosting for compliance is most demanded. What to watch: independent benchmarks on the calibration claims, since “measured, not guaranteed” intervals on unfamiliar data are where extraction models will live or die in production; whether the open-weights release includes the field-type schema as a standard, which would let this become interoperable infrastructure rather than a Cloudflare feature; and whether the extraction-decision-language pipeline pattern gets copied by the cloud rivals now hosting each other’s frontier models, because a three-stage specialized stack undercutting a single LLM by half on real work is the strongest argument yet that the general-purpose model will not be the only thing running in production.