A typed answer can still be wrong. Snowflake and Cloudflare build fast models for decisions and extraction
Snowflake's Decision Fast and Cloudflare's Clairic return categories, scores and values, not prose. They remove parsing failures and cut latency. They do not make a confidence number trustworthy.
Listen to this article · 6 min
AI-generated narration of the full article.
A general-purpose language model that misroutes a support ticket usually fails in a way you can see: it answers in a sentence when you asked for a label, invents a category, or wraps the answer in text your parser rejects. A model built to return a typed decision fails differently. It returns a valid category, with a probability attached, and the category is wrong.
That is the trade two announcements from October 19 make explicit. Both are worth taking seriously, and both shift the work of trusting AI output to a place many teams have not built yet.
Two models that do not write
Snowflake Decision Fast is a lower-latency variant of the Snowflake Decision model introduced two weeks earlier for classification, scoring, filtering and routing. It returns categorical choices with a probability per option, numerical scores and yes or no decisions, instead of free text. It is in private preview, available on request through a Snowflake account team, and reachable in SQL through AI_COMPLETE or from applications and agents through a REST API on Cortex AI Gateway. Snowflake says endpoint round trips in its live demonstration consistently stayed under 300 milliseconds, without describing how that was measured. Pricing is not published.
Cloudflare Clairic, in preview on Workers AI, does for extraction what decision models do for choices. It answers in a single pass, without generating reasoning or output tokens, and charges only for input tokens. It returns extracted spans with character offsets and a confidence score, computed values such as a total from a price plus tax, estimates as a median with an 80% interval, and simple forecasts. When a value is missing it returns null, with the probability that it is null and a best guess. In one of Cloudflare’s examples, a damage claim pipeline ran in 19.3 seconds against 61.7 seconds for an all-LLM pipeline, at $6,032 against $11,740 per million claims. Open weights are planned.
We covered the first wave of this idea when Cloudflare argued that an agent should not ask a frontier model every small question, and when OpenAI started pricing the decision layer separately. The new step is breadth: the same design now covers bounded decisions inside a data warehouse and structured extraction at the network edge.
What typed outputs actually fix
They remove a class of failure that consumes a surprising share of engineering time in production: malformed output. No retries for invalid JSON, no prompt clauses begging the model to answer with one word, no category that does not exist downstream. They also make latency predictable, which matters when a decision sits inside a payment, a fraud check or an agent’s tool selection.
And they make the cost of high-volume classification and extraction look like infrastructure rather than like a chat bill.
What they leave for you to build
Correctness. Cloudflare’s authors say it plainly: outputs are well formed but not necessarily correct. A valid category is a format guarantee, not an accuracy guarantee. Each decision type still needs an evaluation set drawn from your own cases, with error rates by class.
Calibration. Both products expose a confidence number, and both invite you to act on it. Snowflake describes its probabilities as a signal you can threshold on and cites 0.99 as highly confident; it makes no claim about how well those probabilities match real outcomes. Cloudflare says Clairic’s intervals are designed to be calibrated and can be too narrow on unfamiliar data. A threshold is only a control once someone has checked, on your data, that “0.95” is right about 95% of the time. Until then it is a number with a decimal point.
Arithmetic. Clairic computes inside the model’s single pass. Cloudflare reports that one and two step calculations look reliable, while longer dependent calculations and aggregation across many rows are weak, and recommends letting code do longer calculations. That advice generalizes: extract with the model, compute with code, and never let a model estimate stand where a deterministic calculation is possible.
Routing the uncertain cases. The real design object is the policy that sends low-confidence decisions, null values and wide intervals to a person or to a larger model. That is where a cheap model’s errors are caught or escape, which is the cost that matters most, as we argued in pricing models by where their errors land.
The comparison worth running before adopting either product is not cost per call. It is cost per correct decision, with calibration measured, on a few thousand of your own records.