Bottom line

Cloudflare has introduced Clairic, an extraction model available in preview on Workers AI as @cf/cloudflare/clairic. Bypassing traditional autoregressive text generation entirely, Clairic operates over a 65,536-token context window to extract, compute, estimate, and forecast typed values from inputs. The model is built to return structured data alongside confidence scores and probabilities.

Signal

The core signal is the outright elimination of output token generation. Standard large language models simulate data extraction by predicting text tokens that look like JSON. Cloudflare confirms Clairic takes a different approach, natively supporting specifically typed fields: text, number, date, time, datetime, duration, list, estimate, and forecast. Alongside these extractions, the model returns confidence scores and probabilities.

Noise

Applying conventional LLM expectations to Clairic. Because the provider explicitly states it does not generate output tokens, standard generative benchmarks, conversational evaluations, and traditional prompt-engineering methods will not cleanly map to this model. Cloudflare’s "System One-style" designation highlights rapid extraction; it should not be conflated with the complex reasoning capabilities found in conversational models.

Model profile

Provider: Cloudflare

Model: Clairic (@cf/cloudflare/clairic)

Release Date: October 19, 2026 (Preview)

Context Window: 65,536 tokens

Core Capability: Typed extraction, estimation, and forecasting with confidence probabilities.

Where it fits

The operator read is that Clairic fits at the very front of a data ingestion layer. If a pipeline requires taking unstructured text payloads hitting Cloudflare Workers and structuring them before backend processing, this model provides an alternative to using a standard generative LLM. It belongs in deployment scenarios where explicit schema adherence and probability scoring are higher priorities than conversational generation.

Operator implications

The emerging pattern here is workload bifurcation: keeping heavy generative models for reasoning tasks while deploying non-generative, System One-style models at the edge for strict structuring. If the provider facts hold, the likely implication is that operators could potentially reduce data pipeline fragility by shifting basic extraction tasks onto rigid, probability-scored edge primitives that cannot generate stray preamble text.

What is not settled

While Cloudflare states the model returns confidence scores and probabilities with its outputs, it remains unconfirmed if these confidence metadata fields apply uniformly across all supported native field types (e.g., whether a simple "text" extraction carries the exact same scoring mechanism as a "forecast" or "estimate" field). The specific latency profiles and precise extraction accuracy benchmarks compared to standard autoregressive JSON-mode models are also currently unverified in the primary sources.