Not every decision needs a language model

"System One" borrows Daniel Kahneman's split between fast intuitive judgment and slow deliberate reasoning. The model itself is named after William Stanley Jevons, the economist behind the Jevons paradox: make a resource cheaper and total consumption rises rather than falls. TypeSafe is saying out loud what it expects once a decision costs a fraction of a cent.

The architectural claim lands on something visible in nearly every B2B pipeline. Most decisions inside a production system are not System 2 problems. Is this document an invoice or a delivery note. Does this ticket need a human. Is this retrieved passage actually relevant to the question asked. Does this extracted field look complete. Those are classifications, solved with models built to hold a conversation, at conversation latency and conversation prices.

The whole interface is three primitives. Noul answers a yes/no question with a probability between 0 and 1. Choice picks one option from a fixed set of up to 255 entries and returns the full distribution. Score places the case on a scale of two to ten levels and returns a probability-weighted mean.

State and questions share a budget of roughly 64,000 tokens, with state plus the longest single question fitting in about 32,000. Text only, no images, no audio, no video. The narrowness is the point, not a shortcoming: what comes back never has to be parsed, because it was never text.

TypeSafe publishes up to 193 times faster and 444 times cheaper than frontier models on narrow decision tasks. Those are the vendor's benchmarks on tasks the vendor chose. Independent confirmation of the ratio will take months, and narrow benchmark tasks are the friendliest possible case for a decision model.

The direction is not really in question though, and the market made that obvious fast. Vercel added Jev to its AI Gateway and within days called it the fastest adopted model launch in the platform's history. Netlify followed. An open source CRM merged a pull request that introduces an entirely new model category, "evaluation", with a classification node on top of it. Infrastructure does not move that fast for a marginal improvement.

Most AI systems today treat every step as a generation problem, because generation was the only tool in the box. One large model does the routing, the extraction, the judgment and the writing, and the whole pipeline inherits the characteristics of its slowest and most expensive component.

Separate the fast typed decisions from the slow generative work and the shape of the system changes. When a classification costs almost nothing, you stop rationing it. You gate retrieval instead of stuffing context. You score every extracted field instead of sampling. You run the policy check on every request rather than only the ones flagged as risky. Jevons, exactly as advertised.

Jev returns a number without a rationale, which matters the moment someone asks why a case was routed the way it was. It has no world knowledge, so it knows only the state you hand it. Calibration is a property of the distribution across many predictions, not a guarantee about the single answer in front of you. What that means for records and oversight is in A number without a reason.

The sharpest criticism in the Hacker News discussion deserves attention too. Much of the speed may come from constrained single token decoding against a fixed choice set rather than a new architecture. If that holds, the advantage is reproducible and the moat is thin.

None of that changes the design lesson. Decision work and language work have different cost curves, different latency budgets and different failure modes. Building them as one layer was a convenience of the tooling, not a property of the problem. Where the split actually runs through a working pipeline is in Where a decision model actually goes.