TypeSafe AI launches Jev: the model that doesn't write, it decides
TypeSafe AI emerged from two years of stealth with Jev, a model that trades text generation for typed, probabilistic decisions — and claims to run them roughly 100 times faster than frontier LLMs. The numbers are all vendor-reported, and the company is unusually candid about that.
On September 15, a startup called TypeSafe AI ended two years of silence with a launch that cuts against the grain of the frontier-model era: a model that cannot write a single sentence. Its first product, Jev, doesn't generate text — it ingests unstructured context and returns typed, probabilistic decisions a program can act on: a chosen option, a score, a yes-or-no with a confidence attached.
The pitch is simple. Most of what software asks models to do isn't conversation — it's classification, routing, and scoring, repeated millions of times a day through general-purpose models that write a paragraph to deliver a one-word answer. Founder Diogo Almeida, who worked at OpenAI on the instruction-following methods behind ChatGPT, keeps asking why models superhuman at chat have produced so little automation. His answer: the industry built the wrong kind of model for the job.
System One, not chat#
The product category has a deliberate name. TypeSafe calls Jev the first System One Model, after Daniel Kahneman's System 1 — fast, intuitive thinking, as opposed to slow, deliberative System 2. The argument: software has been renting System 2 reasoning to do System 1 chores — routing tickets, qualifying leads, choosing which tool an agent calls next.
Jev is named for William Stanley Jevons of Jevons-paradox fame — cheaper intelligence, TypeSafe argues, expands use rather than shrinking the market.
Under the hood, the company describes a new stack: a new model architecture, a parallel sampler that emits all of a decision's probabilities at once, and a training method called Reinforcement Learning for Calibrated Decisions (RLCD). Because the output is constrained to a predefined schema, TypeSafe says a type error is mathematically impossible — there's no free-text string to parse. Each call supports up to 255 discrete options.
The bet: decisions, not sentences#
Jev entered early access on a waitlist, and the economics are the point: $0.042 per million input tokens, with no charge for output tokens — because there are none. End-to-end responses land between 70 and 500 milliseconds, against seconds to minutes for a frontier LLM on the same task. The thesis: unstructured state in, typed probabilistic decisions out, at a speed and price that make a model call feasible on every hot path.
The demos are chosen to sell that framing: a text-based Doom reading structured game state at about $7 an hour of compute, and a Wikiracing bot navigating hundreds of links per step. Backing is reportedly in place — a $40 million seed round led by DCVC, per industry coverage — though TypeSafe disclosed no funding figures itself.
The numbers — and the asterisks#
This is where the story gets unusual: TypeSafe is strikingly candid about the weakness of its own evidence. The headline figures — roughly two orders of magnitude faster and more efficient, with the home page advertising 193.6x faster and 444.6x cheaper — all come from four workflow evals built by the company's own team. Instead of ground-truth labels, the evals compare Jev against the average predictions of "the largest, smartest, and most expensive" external models — GPT-6 Astra and Fable 5.1 — a reference set TypeSafe itself calls skewed toward OpenAI's and Anthropic's models.
| Claim | Figure | Caveat |
|---|---|---|
| Speed | ~100x faster; 193.6x in one advertised demo | Self-run; demo input shortened, flattering Jev |
| Cost | $0.042/M input tokens, free output; 444.6x cheaper advertised | Company can't yet prove pricing isn't subsidized |
| Latency | 70–500ms end-to-end | Vs. seconds-to-minutes for frontier LLMs on comparable queries (vendor comparison) |
| Hallucination | 0% | A design guarantee from schema matching, not an empirical rate |
| Intelligence parity | "Similar" to existing LLMs on decision tasks | No third-party benchmark; none published |
The "cannot hallucinate" claim deserves care, because it's the easiest to misread. TypeSafe is explicit: the zero figure follows mathematically from enforced schema matching — the model physically cannot emit a value outside the type you defined. That's genuinely useful for pipeline reliability, but it says nothing about whether the decision itself is correct.
Other blanks: how the training data was sourced, public benchmark performance, headcount, founding date. Every comparison runs on TypeSafe's own machines, against comparison models TypeSafe selected — some in non-reasoning mode. Nothing has been independently reproduced yet, though the launch reportedly pulled more than 1,500 Hacker News points within a day.
Why this matters#
The category argument is the interesting part. Agent systems burn most of their latency and budget on the parse-and-hope loop: ask a chat model for a decision, parse the prose, retry when it hallucinates a malformed tool call. A decision engine that makes that loop structurally impossible — not more accurate, but unrepresentable as wrong-shaped output — is aimed squarely at the reliability problem keeping agents out of production.
What to watch#
- Independent reproduction. Everything so far is vendor-graded. The first serious third-party benchmark — someone else's evals, someone else's machines — is the story's real next chapter.
- Pricing durability. TypeSafe is honest it can't prove today's price is sustainable. $0.042 per million input tokens is a bet that inference economics will keep falling.
- What the labs do. If decision-shaped workloads are really a huge share of API traffic, expect the big providers to ship their own structured-decision endpoints rather than cede the segment.
Whether the industry agrees depends on whether anyone outside the company can reproduce the numbers. Watch the benchmarks, not the launch post.