Cloudflare has announced the preview release of Clairic, the first entry in a new class of "System One" models designed specifically for data extraction, computation, estimation, and forecasting. Available immediately on Workers AI as @cf/cloudflare/clairic, the model is hosted directly on Cloudflare’s edge infrastructure, with an open-weights release planned for the near future. According to the company, Clairic handles unstructured inputs—ranging from customer support emails to raw photographs of physical receipts—at significantly higher accuracy, faster latency, and a lower cost profile than traditional Large Language Models (LLMs).
The release follows closely on the heels of the industry-shaking introduction of System One decision models, such as TypeSafe’s Jev and Cloudflare’s own Clef. Grounded in the cognitive framework popularized by Daniel Kahneman’s book Thinking, Fast and Slow, System One models are built to operate swiftly and precisely in a single pass, contrasting sharply with the deliberate, multi-token reasoning characteristic of System Two architectures like frontier LLMs. While decision models like Clef and Jev excel at sorting inputs into predefined categories or calculating choice probabilities, the tech industry has faced a lingering efficiency gap when dealing with the repetitive, structured-output demands placed on traditional AI systems.
Context and the Evolution of System One Models
When TypeSafe introduced the concept of System One models a month prior, it fundamentally shifted how developers thought about integrating non-LLM architectures into automated workflows. Traditional large language models are frequently tasked with routine data-extraction queries: extracting invoice totals, determining delivery dates, calculating shipment delays, estimating repair costs, or forecasting short-term traffic patterns. However, these specific tasks require precise, typed numeric or categorical values rather than the free-form text generation that LLMs are engineered to produce.
Using massive LLMs for structured data extraction often creates a cascade of operational friction. Developers must deploy multi-layered parsing pipelines, schema validators, custom retry mechanisms, and spot-check routines to catch formatting failures, hallucinated digits, or invalid JSON. Furthermore, because LLMs generate text token by token through slow reasoning paths, they incur heavy latency penalties and high costs for every single field across every document processed.
Recognizing these inefficiencies, engineers began exploring alternative model types that could bypass token generation entirely. While previous System One iterations focused exclusively on classification and decision-making, Clairic inaugurates a dedicated category for data extraction and value prediction. Rather than generating open-ended strings, Clairic executes a single parallel forward pass, mapping inputs directly to strongly typed outputs bounded by a predefined schema.
Understanding System One Extraction Models
Clairic is designed to process inputs through a single forward pass without generating tokens, meaning users are charged only for input volume while receiving responses in under a second. This architectural difference eliminates the risk of text-based hallucinations and syntax errors common in LLM outputs. Every response returned by the model is strictly validated against the requested schema type, returning either a precise value or a null result accompanied by rigorous calibration metrics, including a probability of nullity and an explicit confidence score.
The model is capable of parsing diverse input types, including text spans, exact numerical figures, standard calendar dates, times, durations, structured lists, and probabilistic distributions. For estimation and forecasting tasks, Clairic computes median values, confidence intervals, and precise percentile ranges, providing downstream applications with the contextual certainty required to make automated business decisions.
In a modern agentic workflow, these specialized model tiers can operate in unison. Extraction models like Clairic ingest raw, unstructured real-world data and convert it into clean, typed values. Decision models like Clef evaluate those values against established business policies to determine the appropriate course of action. Finally, a traditional LLM can be utilized in the final stage to transform the calculated decision into human-readable prose, ensuring that the generative model is insulated from raw input parsing errors and prompt injection vulnerabilities.
Real-World Applications Across Extraction, Estimation, and Forecasting
To demonstrate the versatility of the new architecture, Cloudflare highlighted several core use cases ranging from customer service automation to time-series forecasting. In document and email extraction scenarios, Clairic processes complex textual inputs alongside optional images to extract specific data fields, such as order identifiers, delivery dates, tax totals, and itemized damage lists. By returning exact character offsets and confidence ratings alongside every extracted value, the system allows developers to verify data provenance instantly.
The model also extends its capabilities to visual inputs, successfully parsing faded or poorly angled photographs of physical receipts to extract line-item totals, purchase timestamps, quantities, and tip amounts. Where traditional optical character recognition or generalized vision-language models might struggle with legibility or formatting inconsistencies, Clairic systematically maps visual data into structured records complete with confidence calibrations.
Beyond basic extraction, Clairic introduces sophisticated estimation capabilities for scenarios where the exact answer is not explicitly written in the source text but can be inferred from surrounding contextual evidence. When tasked with estimating values such as real estate property prices based on comparable sales data, the model returns a structured probability distribution rather than a single unsupported guess. Developers can define custom bounds, logarithmic or linear scales, and step grids, ensuring that the resulting values always conform to strict operational parameters. The width of the returned confidence interval inherently reflects the model’s certainty, giving downstream code the necessary context to determine whether an automated transaction should proceed or be flagged for manual review.
For time-series analysis, Clairic provides forecasting capabilities that bridge the gap between expensive, rigid forecasting pipelines and unreliable LLM prompts. By ingesting historical data provided in standard formats like CSV exports or pasted dashboard tables, the model generates multi-step forecasts complete with confidence intervals. In real-world benchmarks evaluating daily passenger screening numbers at U.S. airport checkpoints provided by the Transportation Security Administration, Clairic successfully captured complex weekly rhythms—such as predictable volume shifts across different days of the week—relying entirely on raw historical counts without requiring prior data cleaning.
Comparative Performance and Workflow Integration
To measure the operational impact of combining these specialized architectures, Cloudflare benchmarked a cascaded pipeline utilizing Clairic for data extraction, Clef for policy-based decision-making, and an LLM for final response generation against a traditional all-LLM pipeline performing the exact same customer support claim workflow.
The performance differential was substantial. The cascaded System One pipeline completed the end-to-end task in 19.3 seconds, compared to 61.7 seconds for the all-LLM approach, representing a 69 percent reduction in execution time. Furthermore, the specialized pipeline reduced overall operating costs by nearly half, driven by significantly lower input and output token consumption. While both pipelines accurately resolved the test cases, the cascaded architecture maintained deterministic consistency across multiple runs, whereas the all-LLM baseline exhibited variable estimation outputs.
Cloudflare has released Clairic as an early preview on Workers AI, allowing developers to integrate the model directly into serverless workers via standard execution calls or REST APIs. While the current endpoint represents an initial checkpoint that will continue to improve over time, the company intends to release the model with open weights in the near future. The preview release is accompanied by comprehensive developer documentation and conceptual guides detailing how extraction models can be integrated into modern application workflows.
