Cloudflare Launches Clairic: Structured Extraction in a Single Forward Pass

Cloudflare has previewed an extraction model called Clairic on Workers AI, positioning it as a "System One extraction model" — one forward pass returns typed values with confidence scores. This description contains two signals worth unpacking: first, inference efficiency is placed at the center of the product design; second, structured information extraction is being peeled away from "generative tasks" and turned into a more fundamental capability closer to classification or regression.

What a Single Forward Pass Means

The conventional approach to information extraction with large models is generation: feed the source text and instructions to the model and have it output JSON or field values. The problem with this approach isn't effectiveness but cost structure and uncertainty — the output must be decoded token by token, and the more fields and the more complex the format, the higher the latency and cost.

Clairic's "System One" naming points to a different approach. In cognitive science, System One refers to fast, intuitive judgments, as opposed to System Two, which is slow and deliberate reasoning. Applied to model design, this means Clairic doesn't rely on multi-step reasoning or chain-of-thought but directly produces results in a single forward pass. According to MarkTechPost, the model returns typed values with confidence scores — that is, each field not only has a value but also comes with a usable reliability metric.

The impact of this combination on engineering deployment is direct. Confidence scores let callers set thresholds: fields above the threshold are adopted directly, while those below are routed to human review or fall back to a stronger model. Extraction tasks shift from "all or nothing" to a pipeline that can be tiered field by field.

Forward Pass vs. Autoregressive Decoding

A forward pass refers to a complete process in which input data flows from the first layer of the network to the last, passing through each layer's computation to produce an output. In generative large models, a forward pass typically produces only one token; to get a complete output, the result must be fed back into the input and called repeatedly — that is, autoregressive decoding. Therefore, the generation length directly determines the number of forward-pass calls, which in turn determines latency and compute cost. When Clairic says "one forward pass," it means the model, in a single call, uses multiple output heads or a structured prediction head to simultaneously produce the values and confidence scores of all target fields, without token-by-token iteration. This is also the fundamental reason output can be free and latency predictable: the amount of computation is determined by input length and fixed output heads, with an approximately linear rather than token-by-token cumulative relationship to the number of fields.

The Engineering Meaning of Confidence Scores

The confidence scores returned by the model usually come from the probability distribution of the output layer (such as softmax probabilities over field values), which, after calibration, are exposed to callers as a reliability metric. Raw probabilities often suffer from overconfidence and require post-processing such as temperature scaling and isotonic regression so that a "0.9 confidence" statistically corresponds to roughly a 90% accuracy rate — a process called calibration. Calibration quality directly determines whether the threshold strategy is reliable: if the score distribution deviates from the true accuracy rate, high-threshold filtering will miss errors, while low thresholds introduce a lot of noise. Therefore, when evaluating Clairic, besides looking at field coverage, you should also pay attention to the calibration curve of its confidence on data from the target domain; this reflects its practical value in a tiered pipeline better than average accuracy alone.

The Pricing Structure Reveals Product Intent

Clairic is priced at $0.10 per million input tokens, with output free of charge. This pricing is quite aggressive in the current AI inference market, but it is not purely subsidized pricing — it is the result of aligning with the "single forward pass" technical approach.

Free output is almost unimaginable for generative models, because output length directly corresponds to computation. Clairic dares not to charge for output, which shows that its output is designed to be short, structured typed values rather than free text. In other words, the billing model itself sends users a message: this model is for extraction, not generation.

For high-frequency extraction scenarios — such as form parsing, log field extraction, and customer service ticket classification — costs can be pushed very low. At one dime per million tokens, many small and medium-sized extraction pipelines have, for the first time, the economic viability to run models directly at the edge.

The Deployment Context of Workers AI

That Clairic chose to preview on Workers AI should not be ignored as a detail of the release channel. Workers AI is Cloudflare's edge inference platform, focused on low latency and tight coupling with the Workers runtime. An extraction model with a single forward pass and short output fits the constraints of edge inference perfectly: controllable memory usage, predictable response times, and no need to maintain long connections.

This also explains why Cloudflare would use "System One" as a selling point rather than a technical footnote. In edge scenarios, users' tolerance for latency is far lower than in cloud batch processing, and the extra round trips brought by multi-step reasoning are often more fatal than the model's own inference time.

Constraints of Edge Inference Platforms

Edge inference means deploying models on network nodes close to users rather than concentrating them in a few large data centers. Cloudflare's global network itself consists of many distributed edge nodes, and Workers AI lets developers run models directly on these nodes. The per-machine compute and memory of edge nodes are far smaller than cloud GPU clusters, unable to host extremely large parameter models or long multi-step reasoning, and they usually adopt a stateless, per-request billing execution model, making it difficult to maintain long connections or sessions. These constraints naturally favor models with moderate parameter counts, completion in a single forward pass, short output, and predictable memory usage. Clairic's design aligns closely with such constraints, which is also an important reason it chose Workers AI rather than a general-purpose cloud inference service as its launch channel.

Divergence in Structured Extraction

Viewed in the larger technological landscape, Clairic represents a divergent path opposite to "use general large models for everything." Over the past two years, information extraction has largely been handed to general models, at the cost of high expense, unstable latency, and output formats requiring repeated validation. The emergence of dedicated extraction models shows that the market is beginning to seek more suitable compute forms for such tasks.

This divergence has already shown signs in the industry. Small dedicated models have long replaced general models in tasks such as classification, embedding, and reranking, and extraction is likely the next stage to be peeled off. Clairic's confidence output further reinforces this trend — it pushes the engineering question of "what to do when the model is uncertain" forward to the model interface level.

Of course, the preview stage means there are still variables undetermined. The calibration quality of confidence scores, the coverage of typed values, and cross-domain generalization all require actual testing to judge. The upper limit of a single forward pass on complex nested structures is also worth observing.

For teams building extraction pipelines, Clairic is worth adding to the evaluation list, especially in scenarios that are latency-sensitive and have high call volumes. But it is too early to treat it as a complete replacement for general models; a more realistic positioning is as the first filter in the pipeline.

One-line Summary

Cloudflare previews Clairic, a dedicated extraction model on Workers AI that returns typed values with confidence in a single forward pass.

Paragraph Summary

Cloudflare has previewed an extraction model called Clairic on its edge inference platform Workers AI, positioned as a "System One extraction model": it does not rely on multi-step reasoning or chain-of-thought but directly returns typed field values with confidence scores in a single forward pass. This design places inference efficiency at the center of the product, making latency and cost predictable. Pricing is $0.10 per million input tokens with free output, and the billing structure corroborates its positioning as short, non-generative output. Confidence output allows extraction pipelines to process field by field: high confidence is adopted directly, while low confidence is routed to human review or a stronger model for verification. The article argues that this represents structured extraction being peeled away from generative tasks and becoming a more fundamental capability closer to classification, forming a divergent path from using general large models for everything.