Cloudflare announced on Tuesday the preview of Clairic, an extraction model that the company describes as the first of a new class of "System One" models designed to pull structured values out of unstructured data without any of the token-by-token generation that defines large language models. The model is available now in preview on Cloudflare's Workers AI platform as @cf/cloudflare/clairic, and the company says it intends to release the model's weights openly in the near future.
The launch positions Cloudflare squarely inside one of the more interesting debates now reshaping the AI infrastructure market: whether every task assigned to an LLM actually needs an LLM. For more context on this story, see our ongoing latest AI developments.
What Clairic Does — and What It Deliberately Doesn't
Clairic is built for a narrow but enormously common job: reading messy, unstructured input and returning typed values. According to Cloudflare's announcement, the model can extract, compute, estimate, and forecast values from unstructured inputs, returning numbers with an 80% range, typed values such as names, dates, times, and lists, and full probability distributions — each with calibrated confidence.
What it does not do is generate text. A query to Clairic returns either a value of the type the developer declared, or a null result paired with a p_null score that estimates the probability the answer simply isn't present in the input. Every answer ships with a confidence score so downstream systems can decide whether to trust it or flag it for human review.
Cloudflare is explicit about the trade-off. Because Clairic generates no tokens, it can't produce malformed JSON, misread a digit mid-generation, or hallucinate an answer in the way an LLM can. It can still make mistakes — the company is careful on that point — but its failures are typed and measurable rather than free-form.
The 'System One' Label, Explained
The terminology comes from the AI startup TypeSafe, which announced its own System One models about a month ago and quickly became one of the most discussed launches in the developer community. The framing borrows from psychologist Daniel Kahneman's book "Thinking, Fast and Slow": System One thinking is fast, automatic, and precise, while System Two thinking is slow and deliberate. TypeSafe mapped that distinction onto AI models, arguing that many production workloads need fast, well-typed answers rather than open-ended reasoning.
TypeSafe's Jev, a decision model purpose-built for a specific domain, took probabilities as inputs and returned probabilities over a fixed list of possible answers. Until now, that was the whole category: every System One model was a decision model.
Clairic extends the concept. Where a decision model picks among predefined options, an extraction model answers with values — numbers pulled directly from the input, computed from it, or forecast beyond it. Cloudflare says the two model families, alongside conventional LLMs, now form a three-layer stack: extraction models turn raw input into typed values, decision models turn values into decisions, and LLMs turn decisions into words.
Why Extraction Is Ripe for a Rethink
Cloudflare's argument rests on how much of today's LLM traffic is spent on work that looks nothing like conversation. The company calls extraction "probably the most common job handed to LLMs today" — pulling an invoice total, a delivery date, or a part number out of a document.
The problem, in Cloudflare's telling, is that LLMs handle that job with machinery built for something else. Output may not be valid JSON. A digit can be misread as tokens are generated. A returned text excerpt can't be traced back to its location in the source document. Teams respond by wrapping models in defensive pipelines: schema validators, retry logic, parsing layers, and spot checks. And because LLMs bill for reasoning and output tokens, every field of every document carries a generation cost that scales with verbosity rather than with the difficulty of the task.
A System One extraction model inverts that economics. Cloudflare says Clairic processes input with a single forward pass, charges only for input tokens, and typically responds in under a second — while a reasoning LLM working through the same fields can take several seconds. For high-volume document processing, the difference compounds quickly.
How a Request Actually Works
Developers send Clairic a state — the text to read, plus up to four optional images — alongside a map of typed fields. Each field carries a name, a type, and optional plain-language instructions. The model returns one typed answer per field, computed in a single parallel pass. If the field list doesn't fit beside the input in one pass, the server splits the work into parallel passes that each see the full input, then merges the answers.
For estimation tasks, answers come as a distribution: a median with an 80% interval spanning the 10th to 90th percentiles, with all 99 percentiles available on request. That makes the model usable for forecasting-style questions — expected repair costs, delivery timelines, traffic over the next day — where a single point estimate would overstate the model's certainty.
A Signal for the Infrastructure Market
Whether Clairic itself succeeds is almost beside the point. The launch is another data point in a fast-moving shift among infrastructure providers toward right-sizing models to tasks: small, fast, cheap models for narrow structured work, with frontier LLMs reserved for genuinely open-ended reasoning. Cloudflare's own positioning — that extraction, decision, and language models should work in unison — reads as a blueprint for how it thinks production AI stacks will be assembled.
The company has not announced a date for the promised open-weights release. For now, Clairic is available in preview on Workers AI, where developers can test it against their own extraction workloads — and find out how much of their LLM bill was ever really necessary.
_This article is based on Cloudflare's official announcement of the Clairic extraction model preview._
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
