Cloudflare Clairic: Structured Data Extraction Without Text Generation

Cloudflare released Clairic on October 19, a Workers AI extraction model (identifier `@cf/cloudflare/clairic`) that returns typed values from a supplied schema instead of generating text, and Cloudflare's own benchmark of a damage-claim pipeline found it 49% cheaper and 69% faster than an all-LLM baseline — $6,032 versus $11,740 per million claims, with latency falling from 61.7 seconds to 19.3. The model accepts up to 65,536 tokens of text plus up to four images and supports eight field types including text, number, date, duration, list, estimate and forecast, returning each answer with a value, a confidence score and a p_null probability. Cloudflare has not published independent validation of the benchmark, and the model cannot explain its answers, limiting it to schema-defined extraction rather than open-ended synthesis.

Cloudflare shipped a model yesterday that does something most AI models refuse to do: it shuts up. Clairic https://blog.cloudflare.com/clairic-extraction-model does not generate text. Point it at an invoice, a support email, a claim form, or a receipt — give it a typed schema — and it returns values: order numbers, totals, dates, durations. Or null if the input does not contain what you asked for. That is the whole job. No prose, no explanation, no hallucinated filler. According to Cloudflare’s benchmark, it is also 49% cheaper and 69% faster than running the same pipeline with a standard LLM. A Model That Does Not Talk Back The LLM world has quietly split into two camps. System Two models — GPT, Claude, Gemini — reason, write, and handle open-ended work. System One models skip token generation entirely: they make a single forward pass over the input and return typed, calibrated answers. No decoding loop, no output buffer, no per-token cost. TypeSafe’s Jev https://typesafe.ai/blog/introducing-system-one-models-and-jev popularized the concept in September. Cloudflare launched Clef https://blog.cloudflare.com/clef-decision-models/ on October 1 for decisions — picking from predefined options, routing support tickets, classifying intent. Clairic, launched October 19, handles the adjacent problem: extracting open-ended typed values from unstructured inputs. There is no list of options to pick from. The answer is somewhere in the document, or it is not. Eight Field Types, One API Call The model identifier is @cf/cloudflare/clairic . You call it like any other Workers AI model, passing a state up to 65,536 tokens of text plus up to four images and a fields map. Supported types: - text — exact spans from the input, with character offsets - number — values read or computed, with optional decimal precision - date / time / datetime — ISO-formatted calendar values - duration — time lengths in specified units - list — multiple matching items - estimate — probability distribution within declared bounds - forecast — up to 64 predicted future steps with 80% confidence intervals Each answer comes back with a value , a confidence score, and a p null probability — the model’s estimate that no answer exists in the input. Text fields also include character offsets so you can locate the span in the original document. const { answers } = await env.AI.run "@cf/cloudflare/clairic", { state: email, fields: { order id: { type: "text", instructions: "Order number" }, amount paid: { type: "number", decimals: 2, unit: "USD" }, delivery date: { type: "date" } } } ; No prompt engineering. No parsing an LLM’s response for a number buried in a paragraph. The response is structured by construction. The Benchmark: 49% Cheaper, But Read the Fine Print Cloudflare ran a damage-claim processing pipeline: Clairic extract then Clef decide then LLM respond versus an all-LLM baseline. Result: $6,032 per million claims versus $11,740. Latency dropped from 61.7 seconds to 19.3. Worth noting: this is a first-party benchmark. Cloudflare designed the test, ran it, and published it. Independent validation has not appeared yet. The cost gap is structurally real — $0.10 per million input tokens with zero output charges is cheaper than any frontier LLM by construction — but treat the specific percentages as directional until third parties verify. When to Use Clairic Instead of an LLM The answer is narrower than Cloudflare’s marketing implies. Clairic wins when you need typed values from a defined schema, not prose; when volume is high and token costs compound; and when latency matters. Stick with an LLM when the output is open-ended, when the task requires synthesis across multiple documents, or when you need an explanation alongside the answer. Clairic cannot explain itself. Also note the documented limitations https://developers.cloudflare.com/changelog/post/2026-10-19-clairic-workers-ai/ : multi-step arithmetic beyond two steps is unreliable, estimate accuracy depends on supporting evidence in the input, and forecasting beyond the immediate next steps flattens toward averages. For anything requiring complex calculations, extract the inputs with Clairic and compute separately. Open Weights: Coming, Eventually Cloudflare says it intends to release Clairic’s weights openly “soon” — no date. Given that Clef shipped open-source on day one Apache 2.0, on Hugging Face , the precedent is there. But Clairic is in preview and Cloudflare mentioned “further checkpoint work” before the open release. Clairic is available now at @cf/cloudflare/clairic through Workers AI, or through AI Gateway for unified control across providers. The pricing is low enough that the question is not whether to try it — it is whether your extraction schema fits the eight field types it supports.