Cloudflare launches Clairic extraction model on Workers AI, open weights planned

Cloudflare announces Clairic, the first System One extraction model

Cloudflare has opened a preview of Clairic, a model hosted on Workers AI under the ID @cf/cloudflare/clairic, with an open-weights release promised soon. Writing on the Cloudflare blog in a post dated 19 October 2026, the company describes Clairic as the first member of a new category it calls System One extraction models: systems designed to pull values out of unstructured input, whether by extraction, computation, estimation or forecasting, and to do so at better accuracy, lower latency and lower cost than traditional LLMs.

Where the idea comes from

The concept builds on work Cloudflare attributes to TypeSafe, which roughly a month before the announcement introduced the notion of System One models, borrowing the label from Daniel Kahneman's book Thinking, Fast and Slow. In that framing, System One is fast and precise, while System Two is slow and deliberate — the territory of LLMs. TypeSafe's Jev, a model purpose-built for a single decision task and priced for its niche, showed that non-LLM models could outperform frontier models on narrow jobs. Cloudflare notes that every System One model so far, including Jev and Clef, has been a decision model: it takes an input state plus a list of questions and returns probabilities over the listed options. Clairic breaks from that pattern by returning values instead of probabilities.

Single pass, no tokens, typed answers

Cloudflare splits most model tasks into three output shapes. Clairic handles typed values such as numbers, dates and lists; Clef handles choices from a fixed set, returning a probability per option; LLMs handle open-ended text. The defining trait of Clairic is that it answers in one parallel pass: it reads the input and computes answers directly, with no reasoning steps and no output tokens. Billing therefore covers only the input, and responses typically arrive in under a second, where a reasoning LLM takes several. Because nothing is generated, Cloudflare argues the model cannot return a malformed string or fabricate an answer the way an LLM can — though it can still be wrong. Every answer is either a value of the declared type or null, and each carries a confidence score plus p_null, the probability that the value simply is not present in the input, giving callers a signal for when a second look is warranted.

What it can return

According to the blog post, Clairic produces four kinds of results:

• Extracted values: spans copied straight from the input, such as a name or an order ID, returned with character offsets and a confidence score so each answer traces back to its source.
• Computed values: numbers, dates, times, durations and lists that are read from the input or derived from it, obeying the rules of their type.
• Estimated values: distributions for quantities the input implies but never states, confined to bounds and a step grid the caller declares, reported with a median and an 80 percent interval.
• Forecasted values: extensions of a series into the future, one estimate per step, each carrying its own 80 percent interval.

How a request works

A request consists of a state — the text to read, plus up to four optional images — and a map of typed fields. The caller names each field, picks a type such as text, number, date, datetime, duration, list, estimate or forecast, and can attach plain-language instructions. The model returns one typed answer per field from a single forward pass. There is no cap on the number of fields: if they do not fit alongside the input in one pass, the server splits them into parallel passes that each see the full input and merges the results. Cloudflare contrasts this with typical LLM extraction workflows — asking for an invoice number, vendor and due date as JSON — which often need stacks of parsers, schema checks, retries and manual review to catch invalid output or misread digits.

Why it matters

A large share of the structured-output calls hitting frontier LLMs today are simple field lookups: invoice totals, delivery dates, days late. Cloudflare's pitch is that these jobs do not need token-by-token generation or chain-of-thought reasoning, and paying for reasoning and output tokens on every field of every document is waste. If Clairic's accuracy and latency claims hold up in practice — and the open-weights release actually lands — it sketches a division of labour in which extraction models turn raw input into typed values, decision models turn values into decisions, and LLMs are reserved for writing and reasoning. For teams running document pipelines at scale, that could mean cheaper and more verifiable extraction with fewer silent failure modes.