Cloudflare has a new idea about what you are overpaying for. On October 19, the company launched a preview of Clairic, an AI model that does one very specific job: read unstructured text and images, and return typed values — totals, dates, lists, estimates, forecasts — in a single forward pass, without generating a single output token.
Cloudflare calls Clairic the first model in a new "System One" extraction class, borrowing the Kahneman terminology popularized in AI by TypeSafe's Jev decision model. The pitch is that developers routinely hand jobs like "read this invoice, return the order number, total and due date as JSON" to frontier large language models, then pay output tokens and stack parsers and retry logic around the failure modes: broken JSON, a misread digit, an excerpt with no pointer back to the source. Clairic answers with values directly — extracted, computed, estimated, or forecast — each arriving as a typed result with confidence scores, or null when the input simply does not contain the answer.
The model that returns values, not text
You send Clairic text (up to four embedded images can ride along) plus a map of named, typed fields, and it answers each one in a single pass. Extracted text comes back with character offsets. Computed values — a due date derived from "net 30" terms, a total summed from line items — follow the rules of their type. Estimates come as a median with an 80% interval inside bounds you declare. The model context holds 65,536 tokens, and fields that overflow the 8,192-token pass are split across parallel passes.
Pricing is built around the zero-output-token design: Cloudflare charges only for input, at $0.10 per million input tokens. The model is available in preview on Workers AI as @cf/cloudflare/clairic, callable from a Cloudflare Worker or through the REST API. Cloudflare says it intends to release the model weights openly after further checkpoint work, though it has not named a date. For now, it runs only on Cloudflare's endpoint, and it ships with an "AS IS" caveat — Cloudflare can stop offering it without a refund.
49% cheaper, 69% faster — in Cloudflare's own test
In Cloudflare's internal claims-pipeline test, a pipeline pairing Clairic for extraction with Clef — the company's System One decision model — finished in 19.3 seconds against 61.7 seconds for an all-LLM pipeline, and cost 49% less. On a public TSA daily-screening dataset, Clairic's forecasts missed by an average of 3.3%, with 17 of 21 actual counts landing inside the predicted 80% interval. Cloudflare is candid about the limits: multi-step arithmetic, calibration on unfamiliar data, and long-range forecasting are weak spots, and a well-formed answer is not necessarily a correct one.
The launch fits a pattern of Cloudflare building specialized, single-job models instead of generalist LLMs — Clef handles decisions, Clairic handles extraction, and the general model only steps in when neither can answer. If the benchmarks hold up in real workflows, the biggest shift may be economic: structured-output calls are among the most common workloads sent to frontier models, and a model that prices them by input alone could reset what developers expect to pay for routine extraction.
By Hamza Faisal
