Chat models write strings. System One returns typed decisions.

Chat stacks optimize for text people can read. Teams then wrap those replies in parsers, JSON mode, and retries so software can act. TypeSafe’s pitch is that the wrapper still cannot buy what System One is built for. Calibration (how well stated confidence matches real error rates), type-safety as design rather than a post-hoc schema, and parallel sampling of decisions instead of sequential token sampling for a reply.

The training method named on primary pages is Reinforcement Learning for Calibrated Decisions (RLCD), alongside a new architecture and a parallel sampler. The company contrasts that stack with RLHF and with verifiable-reward chat setups. Take that as company framing. It is the bet, not an independent proof.

The naming is deliberate. “System One” borrows Kahneman’s System 1 / System 2 language (fast, automatic judgment versus slow deliberation). “Jev” points at William Stanley Jevons, the economist linked to the idea that efficiency can raise demand rather than cut it. TypeSafe puts those labels in its FAQ. They are branding with an argument inside them. Cheaper, faster decisions may mean more automated decisions, not fewer.

Docs describe three question primitives evaluated in parallel against one state. Choice picks among options and returns confidence. Score rates something and returns confidence. Noul is the third primitive in that set. The HTTP surface is POST /v1/systemone. The default alias is jev-latest.

In plain terms, the model is not writing you an essay. You hand it state and typed questions. You get typed answers you can compose in application code. That is what “decisions composed in code” means here. The unit of work is a decision object, not a chat turn.

Access is early access and waitlist. Jev accepts text today (strings, JSON, text arrays). Images, audio, and video are not supported yet on the docs.

Company pages list published prices. Input is $0.042 per million tokens ($42 per billion input tokens). Output tokens are described as free or too cheap to meter on the announce comparison table. The homepage markets “193.6x Faster, 444.6x Cheaper” on System One workflow evals. The blog carries caveats (including West Coast laptop latency and workflow evals built by their capabilities team, with a reference average of large chat models the blog names as Astra and Fable 5.1).

Every speed and cost multiple above is a dated vendor claim with those nuances. Homepage multiples are company marketing, not independent measurements. Independent Arena, MMLU, or GPQA scores for Jev do not appear in the cited primary sources.

TypeSafe’s stated eval stance matters for how to read any number they publish. The company says it does not ship a standard public-benchmark table in releases. It runs new evals as dated snapshots and then retires them. Workflow evals compare models inside fixed code graphs against a reference of large chat models. The antibenchmaxxing essay is the company’s argument for that stance.

That is a trust move as much as a method note. If “how smart is the chat model?” is the wrong question for typed automation, then a chat leaderboard is the wrong trophy case. Whether TypeSafe’s own workflow evals are fair is a separate question. Label them as company method. Do not launder them into third-party consensus.

Docs also undercut absolute safety marketing. Calibration does not guarantee that any single answer is correct. Schema and type constraints are design claims. They are not a promise of zero errors. Keep company language in that register.

The working thesis is simple. Chat models optimize for human-readable strings. System One and Jev bet that reliable automation needs typed, calibrated decisions composed in code. So the right question is not which chat model sounds smartest. It is whether a decision object, with a type and a confidence, is trustworthy enough to drive the next line of software without a human in the loop for every turn.

TypeSafe has shipped the first public System One model under that bet. Early access. Text in. Typed decisions out. Vendor speed and cost claims on the table, with blog caveats attached. Independent public-benchmark scorecards are not in the cited company materials.

If you build software that has to choose, score, or refuse, the product split is the news. The chat scoreboard was always a proxy. Jev is an argument that the proxy is no longer the product.