TypeSafe launches Jev, a model for structured decisions, not chat

TypeSafe launched its first System One Model, called Jev, into early access on September 15, 2026, after two years in stealth. The announcement comes from founder Diogo Almeida, who describes working at OpenAI on the instruction-following methods that became the research behind ChatGPT. Almeida frames the launch around a question described in the post as the driving one for the last four years: language models have been superhuman at chat for years, so why has automation not kept pace? The answer, in Almeida's telling, is Jev, a new class of model built specifically to make fast, structured decisions that software can consume directly, rather than to generate conversational text.

Jev runs on what TypeSafe describes as an entirely new stack: a new model architecture, a parallel sampler the company says maximizes efficiency, and a training method it calls Reinforcement Learning for Calibrated Decisions (RLCD). Unlike a large language model, which generates output token by token, Jev outputs all of its probabilities in parallel and gives up string generation entirely; TypeSafe describes it as 'a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.' The company says Jev reaches similar levels of intelligence to existing LLMs on these structured-decision tasks while running about two orders of magnitude (roughly 100 times) faster and more efficient.

To back the speed claim, TypeSafe built what it calls workflow evals: rather than scoring models against a fixed ground truth, the evaluation assumes a workflow can be represented as a compute graph in code and uses the predictions of 'the largest, smartest, and most expensive' external models as a reference, specifically the average of two models TypeSafe names GPT-6 Astra and Fable 5.1. On that basis, TypeSafe says Jev owns the Pareto frontier by almost two orders of magnitude across four published workflows (the axis, whether speed, cost or accuracy, is not specified), and that the 193.6x faster and 444.6x cheaper figures advertised on its home page come from the high end of those same four workflows.

TypeSafe pairs the speed claim with a type-safety one: because Jev's output is constrained to match a schema, the company says a type error is mathematically impossible, and it plots Jev's hallucination and type-error rate at 0 percent. TypeSafe is explicit that this figure is not an empirical measurement but a guarantee that follows from schema matching. It separately notes that the comparison figures it uses for the LLMs come from OpenRouter, a source it says almost certainly carries bias, since harder queries may be routed to stronger models; the post states only that existing models still hallucinate and have type errors, without giving a comparable measured rate for them.

TypeSafe also published two demos. In one, a text-based version of Doom runs on Jev reading structured game state rather than images; an engineer on the project worried that running 10 queries a second would be too costly, but the team found the resulting cost, roughly $7 an hour, lower than expected, and TypeSafe plans an in-depth walkthrough plus live events built around it. In the second, Jev plays Wikiracing, navigating from one Wikipedia page to another using only links encountered along the way, with each step offering a choice of hundreds to thousands of links. TypeSafe says Jev supports a decision cardinality of up to 255 options, scoring choices independently and then making an explicit pick for the largest option sets, which causes occasional slowdowns. The company says its speed advantage is smaller here than in the other demos because the comparison LLMs ran in non-reasoning mode (with one exception, Astra, run at its lowest reasoning setting) to keep the demo watchable, and that Jev also finished in fewer steps, which TypeSafe calls a sign of greater intelligence.

Through the post, TypeSafe repeatedly flags its own evidence as biased. Its side-by-side demo uses a simplified query with a shorter input than a typical prompt, which the company admits 'paints our model in an advantageous light,' and it picked GPT-5.6 Terra, run with default reasoning, as the comparison because TypeSafe judged it the most comparable to Jev in intelligence. The reference score in the workflow evals, the average of GPT-6 Astra and Fable 5.1, is described by TypeSafe as skewed toward OpenAI's and Anthropic's models, which the company says likely makes Jev's relative performance against DeepSeek's models look worse than it actually is. TypeSafe also notes the workflow content itself was built by its own model-capabilities team, so, in its own words, 'some bias could exist,' even though the content was not deliberately constructed to flatter Jev and is not part of its training data.

TypeSafe explains both names in an FAQ at the end of the post: System One Models nods to Daniel Kahneman's distinction, from Thinking, Fast and Slow, between fast intuitive System 1 thinking and slow deliberate System 2 reasoning, and Jev is named for the economist William Stanley Jevons, whose work TypeSafe cites as the reason it expects cheaper intelligence to expand use cases rather than shrink the market for them. Most of the FAQ's other questions, including how Jev's training data was sourced, how it performs on public benchmarks, and whether it is simply a smaller LLM, are left unanswered in the retrieved page. The post also gives no per-call dollar price despite calling its pricing transparent, and discloses no funding amount, valuation, investor names, headcount or founding date for the company.

Key facts

TypeSafe, founded by Diogo Almeida (previously at OpenAI, per the post), opened early access to Jev on September 15, 2026, after two years in stealth.
Jev is TypeSafe's first System One Model: it outputs typed, probabilistic decisions in parallel instead of generating text, built on a new architecture, a parallel sampler, and a training method TypeSafe calls Reinforcement Learning for Calibrated Decisions (RLCD).
TypeSafe claims Jev matches existing LLMs' intelligence on these tasks while running about two orders of magnitude (roughly 100 times) faster and more efficient, with home-page figures of 193.6x faster and 444.6x cheaper drawn from the high end of four published workflow evals.
Because Jev's output must match a schema, TypeSafe says a type error is mathematically impossible and plots the model's hallucination rate at 0 percent, though the company itself calls that figure a guarantee by design rather than an empirical measurement.
TypeSafe discloses its own evaluation biases: the workflow-eval reference score averages only GPT-6 Astra and Fable 5.1, skewing toward OpenAI's and Anthropic's models, and the comparison LLMs' hallucination numbers are sourced from OpenRouter, which the company says almost certainly carries routing bias.

Why it matters

TypeSafe is proposing a new product category, not a faster chat model: a System One class built specifically to make decisions that software can consume directly, positioned as the missing automation layer beneath chat-capable LLMs. The pitch is that giving up open-ended text generation buys speed, cost and guaranteed schema conformity, at claimed gains of roughly two orders of magnitude, for the narrower job of structured decisions rather than open conversation. The company frames this historically too: it names Jev for the economist William Stanley Jevons, arguing that cheaper intelligence expands use cases rather than shrinking the market for it, much as cheaper steam power increased coal demand rather than reducing it.

Who it affects

The immediate audience is developers building automation that currently routes a decision through an LLM: TypeSafe is moving people off an early-access waitlist and explicitly asking which decisions they want to automate. The company frames the stakes around systems with latency guarantees or agents with deeply nested dependency chains, where it says a hallucinated tool call is 'an absolute deal-breaker' rather than an inconvenience. Indirectly, the story challenges general-purpose LLM providers: TypeSafe names GPT-6 Astra, Fable 5.1 and GPT-5.6 Terra as its points of comparison, and says its main workflow-eval reference, the GPT-6 Astra and Fable 5.1 average, is skewed toward OpenAI's and Anthropic's models specifically, likely understating how Jev compares against DeepSeek's, which were not part of that reference set.

How to use it

Jev is live in early access; TypeSafe says it is moving developers off its waitlist 'as quickly as we can' and wants feedback on which decisions people need automated and where Jev falls short. The interface is a single call, unstructured state in, typed decisions out, and the Wikiracing demo shows it handling as many as 255 discrete options in one decision, scoring larger option sets in two stages to control latency. TypeSafe describes its pricing as transparent and expects it to fall over time, but the post gives no per-call dollar figure, and the company says outright that it cannot yet prove its current pricing is not subsidized.

How solid is it

Every figure here is TypeSafe's own, and the post is unusually candid about where its evidence is weak. The published speed evals run from TypeSafe's own West Coast machines. The side-by-side demo uses a shortened input that the company admits 'paints our model in an advantageous light,' and the comparison model in that demo, GPT-5.6 Terra run with default reasoning, was chosen by TypeSafe itself as the one it judged most comparable to Jev. The workflow-eval reference, the average of GPT-6 Astra and Fable 5.1, was also chosen by TypeSafe, which says it biases results toward OpenAI's and Anthropic's models; the workflow content itself, while said not to be built to flatter Jev, was authored by TypeSafe's own model-capabilities team. The headline 0 percent hallucination figure is not measured: it is a guarantee that follows from enforced schema matching, and the comparison figures for rival LLMs come from OpenRouter, which TypeSafe itself calls likely biased. No named third-party benchmark backs the claim that Jev matches existing LLMs on intelligence, and TypeSafe does not say which axis, speed, cost or accuracy, its Pareto-frontier claim is measured on. Most of the post's own FAQ, including how the training data was sourced and how Jev performs on public benchmarks, is unanswered in the retrieved page.

Risks and caveats

Jev cannot generate open-ended text. It is not a chat or writing replacement. TypeSafe's claim that Jev cannot hallucinate covers structural conformity to a schema, not whether a given decision is correct: the company's own side-by-side demo shows Jev and GPT-5.6 Terra disagreeing on one output, 'Churn likelihood level,' which TypeSafe itself calls genuinely ambiguous. The announcement discloses no per-call price, no funding amount or valuation, no investor names, no headcount and no founding date for the company. It is a launch post, not an independent evaluation, and none of the comparisons here have been reproduced outside TypeSafe.

“Models have been superhuman at chat for years, so where is all the automation?”

— Diogo Almeida, founder of TypeSafe, in the Jev launch post