TypeSafe AI

Jev

Provider

TypeSafe AI

Status

Preview

Context

64,000 tok

Price

$0.042 / $0 /MTok

Jev is the first model from TypeSafe AI, announced on 15 September 2026, and it is built on a different idea from every other model on this site. It does not generate text. You send it some state, such as a support ticket, a document or a game’s current position, plus a set of typed questions, and it returns typed decisions: a choice from a fixed list, a score, or the probability that the answer is yes. Every answer carries a confidence estimate your code can act on (TypeSafe).

TypeSafe calls this a System One model. The target is the kind of software decision that is often handed to a chat model today and parsed back out of its prose: routing a request, classifying a document, scoring a lead, checking a guardrail. For those jobs TypeSafe argues a model should answer in the right type, at once, and say how sure it is.

Quick specs

How Jev works

Jev answers three kinds of question, which TypeSafe calls primitives (TypeSafe docs):

• Choice picks from a list of options you define and returns the choice, the probability of each option and a confidence figure. TypeSafe caps the number of options at 255.
• Score rates the state against criteria you set and returns the score, its probabilities and a confidence figure.
• Noul asks a yes/no question and returns a single number from 0 to 1, the probability that the answer is yes (TypeSafe docs).

All three can be mixed in one API call and are evaluated in parallel. The docs advise asking each question about one specific, well-scoped thing and combining the answers in your own code, rather than asking one broad question.

Two design choices separate Jev from a language model. It produces its answers in a single parallel pass rather than one token at a time, which TypeSafe credits for its speed. And TypeSafe trains it with Reinforcement Learning for Calibrated Decisions (RLCD), which rewards honest probabilities rather than answers people prefer. TypeSafe says Jev is calibrated, so higher confidence means higher accuracy, and its docs are careful to add that confidence is measured across groups of predictions, not a guarantee for any single answer.

What TypeSafe claims, and where the numbers come from

TypeSafe’s launch figures are large, and it publishes the method behind them, which lets you read them properly.

Three caveats come from TypeSafe itself. The workflows behind the multiples were built by its own model capabilities team, and it says “some bias could exist”. It describes the results as the higher end of real-world gains. And its zero-hallucination claim rests on the fact that Jev’s output always matches the requested type, so it cannot return a malformed answer. TypeSafe says of that figure, “our number is not empirical” (TypeSafe). Jev can still pick the wrong option, and the confidence score exists to flag when that is likely.

Independent coverage adds two more. SiliconANGLE notes the performance claims “have not been independently verified” and will vary by workload, network location and comparison method (SiliconANGLE). The Register called the hallucination comparison unfair, because Jev’s output is not natural language. The price comparison has the same limit: an LLM bills for the text it writes, while Jev’s output is free and all of its cost sits in the input.

The Register also reported a launch demo of Jev playing Doom from structured game state, a fair illustration of the real-time loops TypeSafe is targeting.

Pricing and access

Jev costs $0.042 per million input tokens, or $42 per billion, and output is free (TypeSafe docs). SiliconANGLE reports a cost of $0.39 per 1,000 workflows (SiliconANGLE). It is available in early access: developers join a waitlist and are admitted from it, then call it through the HTTP API or the Python and JavaScript SDKs, with keys from TypeSafe’s console. The weights are not released.

What Jev is for, and what it is not

TypeSafe’s cookbooks show the intended jobs: intent routing, classifying retrieval passages, re-ranking search results, checking citations, guardrails for LLM output, date and value extraction, hierarchical classification and choosing which function to call (TypeSafe docs). The common thread is a high volume of repeated decisions where the possible answers are known in advance.

It is not a replacement for a chat model. Jev does not write text, code or explanations, and TypeSafe’s own patterns pair it with LLMs rather than replacing them. One of them uses Jev’s confidence to decide when to escalate a decision to a reasoning model or a person.

Known limitations

TypeSafe publishes an unusually direct list of where Jev 1.13 goes wrong (TypeSafe docs):

• Literal reading. It answers the question as written, not as intended.
• Maths and numbers. TypeSafe says “Jev is not a calculator”. It does not count reliably or judge whether two values are close.
• Dates and times. It reads dates as text, not as ordered quantities.
• Indirection. Double negatives and nested instructions are answered less reliably.
• Irrelevant state. Accuracy falls as the input fills with content unrelated to the decision.
• Adversarial content. Text written to steer the model can move the answer.
• Conflicting instructions. When instructions and criteria disagree, the answer can suffer.
• Consistency across questions. Related questions are not guaranteed to produce mathematically consistent answers.

Add to that the product limits: text input only, with no images, audio or video yet, a 64k-token request cap and a maximum of 255 options per choice.

Why Jev is not on our ranking

Our ranking of AI models places general-purpose models on independent evidence, led by the Artificial Analysis Intelligence Index. Jev has no score there and cannot have one, because the index measures tasks that require generating text. TypeSafe publishes no results on standard public benchmarks, and its own comparisons come from workflows it built. Ranking it against GPT-6 Astra or Claude Fable 5.1 would compare different kinds of tool, so Jev is covered here on its own terms instead.

Frequently asked questions

What is Jev?

Jev is a System One model from TypeSafe AI, announced on 15 September 2026. Instead of generating text, it reads text state and returns typed decisions, such as a choice, a score or a yes/no probability, each with a calibrated confidence estimate that software can act on directly.

How is Jev different from ChatGPT or Claude?

ChatGPT and Claude generate text one token at a time and are trained on human preferences. Jev returns typed answers in a single parallel pass and is trained with Reinforcement Learning for Calibrated Decisions, which rewards honest probabilities. It cannot write text, code or explanations, so it complements a chat model rather than replacing one.

How much does Jev cost?

Jev costs $0.042 per million input tokens, or $42 per billion, and output tokens are free. It is in early access, so developers must join TypeSafe’s waitlist first.

Is Jev really hallucination-free?

Only in a narrow sense. Jev’s answers always match the type you ask for, so it cannot return a malformed or invented output format. TypeSafe itself says its zero-hallucination figure “is not empirical”. Jev can still choose the wrong answer, which is what its confidence score is there to flag.

How fast is Jev?

TypeSafe reports end-to-end response times of 70 to 500 milliseconds, against 3 to 329 seconds for frontier LLMs on its own workflows. Those figures have not been independently verified, and TypeSafe says its workflows may carry some bias.

Is Jev open source?

No. Jev is a proprietary model available through TypeSafe’s API, and its weights and parameter count are not published.

Why isn’t Jev on the AI models ranking?

Our ranking scores general-purpose models on independent benchmarks such as the Artificial Analysis Intelligence Index. Jev does not generate text, so it cannot be scored on those benchmarks, and TypeSafe publishes no standard benchmark results. Comparing it with general LLMs on the same board would mislead.