Cloudflare Launches Clairic: System-One Extraction Model with Zero Token Output Fees
Cloudflare has debuted Clairic on Workers AI, a non-autoregressive 'System One' data extraction model returning strongly typed values directly in a single forward pass without token streaming latency, priced at $0.10/M input tokens and zero output fees.
Cloudflare has officially launched 'Clairic', an innovative machine learning architecture deployed on Workers AI (@cf/cloudflare/clairic) engineered specifically for structured extraction from unstructured enterprise data. Grounded in Daniel Kahneman's cognitive framework of 'System 1'—rapid, intuitive pattern recognition absent protracted deliberation—Clairic resolves extraction queries in a single forward pass without autoregressive token generation.
Historically, engineering teams routed unstructured invoices, shipping manifests, and contracts through general-purpose Large Language Models. However, conversational LLMs introduce high per-token output pricing, multi-second decoding latency, and structural hallucination risks. Clairic fundamentally reimagines this workflow by returning verified, typed data payloads directly to client applications.
Extraction Mechanics: Strongly Typed Schemas and Zero-Hallucination via p_null
Developers supply document text (and up to four images) alongside a typed schema specification. Operating across a 65,536-token context window, Clairic resolves fields into standardized native types:
- text: Verbatim document substrings accompanied by exact character-offset boundary pointers.
- number, date, time, datetime, duration: Standardized numerical and chronological values normalized to ISO formats.
- estimate & forecast: Statistical distributions delivering median values with 80% confidence intervals, alongside forward projections up to 64 steps.
- p_null (Omission Probability): Every extracted field returns confidence metrics paired with a p_null probability score. When a requested field is absent from the source, the model returns a deterministic null rather than generating plausible hallucinations.
Serverless Unit Economics: $0.10 per 1M Input Tokens with Zero Output Costs
By eliminating iterative next-token generation loops, Clairic operates sub-second inference across Cloudflare's global edge network. Cloudflare introduced a disruptive pricing structure: US$0.10 per 1 million input tokens.
Crucially, output token charges are eliminated entirely (US$0.00). Because data is mapped directly into binary payload representations rather than rendered as streaming strings, enterprise data extraction workloads achieve up to a 90% cost reduction compared to legacy frontier LLM API pipelines.
Architectural Specialization and System Boundaries
Cloudflare explicitly delineates the operational boundaries of System One architecture: while the model reliably computes single-step arithmetic, protracted mathematical derivations should remain isolated within deterministic application code.
- First non-autoregressive extraction foundation model on Cloudflare Workers AI (@cf/cloudflare/clairic)
- Features a 65,536-token context window and multimodal ingestion for up to four document images
- Outputs strongly typed payloads: verbatim text with offsets, numbers, datetimes, durations, and statistical distributions
- Hallucination mitigation via rigorous confidence scores and explicit p_null omission detection
- Flat pricing of US$0.10 per 1M input tokens with zero output token fees
- Open model weights slated for independent distribution in an upcoming release
“The evolution of production AI requires decoupling tasks: using fast, non-autoregressive models to extract ground truth facts in milliseconds, reserving heavyweight reasoning models strictly for creative narrative generation.”
Clairic exemplifies a broader industry transition toward composable, single-purpose AI architectures that deliver sub-second determinism and dramatic operational cost reductions directly at the network edge.
References & Primary Sources
Related Analysis & Context
Modal Adopts Nvidia Vera CPUs for AI Agents: Doubling Execution Throughput in Serverless Cloud
Serverless AI cloud platform Modal has selected Nvidia's new Vera CPUs to power agentic infrastructure, doubling task throughput compared to incumbent x86 chips amid 15x YoY revenue expansion toward a $15.75B valuation.
Cloudflare Gateway: Inline Workers Proxy & Programmable DLP
Cloudflare has integrated Cloudflare Gateway and Data Loss Prevention (DLP) directly with Cloudflare Workers across 330+ edge locations, enabling zero-hairpinning inline HTTP mutation and live programmatic DLP secret verification.
Cloudflare & OpenAI: Real-Time Signals for Web Discovery
A joint engineering pilot across 1% of Cloudflare's global network enabled OpenAI to discover hundreds of millions of previously unindexed web pages within seconds of publication while cutting redundant crawler polling and 404 traffic.