Cloudflare introduces Clairic: Clairic is a new model from Cloudflare that

Cloudflare launched a preview of Clairic, a specialized System One extraction model designed to pull, compute, estimate, and forecast structured values from unstructured inputs. Hosted on Workers AI under the model identifier @cf/cloudflare/clairic, the system processes tasks in a single parallel pass without token-by-token generation, delivering responses in under a second while running at higher accuracy and lower operational cost than traditional large language models, as outlined in Cloudflare Blog's report. The company confirmed plans to release open weights for the model in an upcoming distribution phase.

This deep dive examines the operational architecture of System One extraction models, detailing how typed field definitions, character offset spans, and bounded numerical ranges replace token generation pipelines. It is written for backend engineers, data pipeline architects, and infrastructure teams seeking deterministic structured extraction from noisy documents, receipts, support tickets, and time series data without paying for generative output tokens or reasoning passes.

01Single forward pass and typed value extraction

Clairic processes inputs through a single forward pass over input tokens rather than decoding text token by token. Because it does not run iterative generation cycles, the model charges exclusively for input tokens and eliminates output token fees entirely. It guarantees outputs that match declared schemas, returning typed values or null when an answer cannot be located in the provided context.

The system bypasses the primary failure modes of frontier language models, which often emit unparseable JSON, drop digits during token generation, or require external validation harnesses. By reading structured fields directly, Clairic delivers character offset spans for text spans, exact decimal formatting for currency or metrics, and ISO formatting for dates, times, and datetimes.

Every scalar response carries a confidence score alongside p_null, which calculates the exact probability that the requested value does not exist in the source text. When fields cannot fit beside the input context in one cycle, the Workers AI server automatically shards them across parallel passes that inspect the full input and merges the returned fields into a unified schema.

02Functional split across extraction, decision, and generation

Clairic expands the System One paradigm popularized by TypeSafe, which introduced Jev a month prior to handle probabilistic categorization against structured options. While decision models like Jev and Clef evaluate finite candidate choices, Clairic establishes an extraction category that computes continuous numbers, extracts text spans, and forecasts future data points.

System One decision models classify inputs into predefined categories to help applications determine discrete actions. System One extraction models sit ahead of that layer by converting raw, unstructured inputs into strongly typed fields, numbers, and distributions. General-purpose System Two models like GPT, Claude, Gemini, and DeepSeek remain reserved for deliberate reasoning, multi-step synthesis, and open-ended text composition.

Within an enterprise pipeline, these architectures operate in sequence to ingest and execute workflows. Extraction models first ingest unstructured records to output typed values, decision models ingest those values to determine appropriate actions, and generative models transform final decisions into conversational prose or customer-facing correspondence.

03Four output classes across spans and distributions

Requests sent to Clairic define a state containing input text alongside up to four optional images, paired with a schema map of typed field declarations. The extraction model supports four output categories: extracted text values, computed values, estimated ranges, and multi-step forecasted series.

Extracted values return verbatim strings such as customer names, tracking codes, or order identifiers paired with exact start and end character offsets. Computed values evaluate durations, timestamps, and numbers, adhering to strict typing constraints like two decimal places for cents or explicit temporal lengths for durations.

Estimated values handle unstated properties supported by contextual evidence, returning distributions bounded by user-defined minimums, maximums, and grid steps with median markers and 80 percent intervals from p10 to p90. Forecasted values project future intervals sequentially, assigning an 80 percent interval to each subsequent step in a series.

04Processing multi-field payloads on Workers AI

Engineering teams can query Clairic directly within Cloudflare Workers through the Workers AI binding using the @cf/cloudflare/clairic endpoint. Developers construct field requests with plain-language guidance, data type definitions, and optional measurement bounds without needing prompt engineering scaffolds or format repair layers.

Because field evaluations compute concurrently in parallel passes, batching multiple field extractions introduces minimal latency overhead compared to querying a single attribute. Requesting thirty distinct fields from a multi-paragraph support ticket costs roughly the same and executes in the same single-second envelope as requesting three fields.

Pipelines that previously maintained multiple layers of schema validation, retry logic, and JSON parsing can route unstructured text directly into Clairic. In customer service examples, such as an email detailing order numbers, purchase amounts, taxes, and arrival dates, the model maps invoice sums, dates, and names into verified variables in one request.

05Production migration from frontier reasoning models

Adopting System One extraction models reduces reliance on frontier language models for high-volume, structured extraction workloads. Queries evaluating delivery dates, repair bills, invoice totals, and daily traffic estimates shift away from token-heavy reasoning engines toward deterministic regression and span extraction models.

Applications replacing token generation with single-pass inference eliminate runtime hallucination risks, as values must either map directly to input spans or satisfy strict mathematical and temporal types. If a requested value is missing from the payload, the model outputs null alongside calibrated uncertainty metrics rather than inventing plausible figures.

Teams preparing workflows for the model can deploy against the preview endpoint on Workers AI while monitoring upcoming open-weight releases. Integrating extraction models at the intake layer establishes low-latency pipelines that feed downstream decision engines and agentic workflows without generative token overhead.