01
Overview
Jev is a new kind of AI model from TypeSafe AI built to make decisions inside software rather than write prose for people. TypeSafe calls the category a System One Model. An application sends Jev some state plus predefined questions; Jev returns typed answers with probabilities and confidence information that code can use directly.
That makes Jev less like a chatbot and more like an intelligence primitive for routing, classification, scoring, gating and agent control.
02
What is Jev?
TypeSafe AI introduced Jev on September 15, 2026 as its first public System One Model. The company says it built a new architecture, parallel sampler and training method called Reinforcement Learning for Calibrated Decisions (RLCD).
Unlike an LLM that generates a sequence of text tokens, Jev is designed around structured decisions. Its public API supports decision types including choices, scores and yes/no-style probability judgments. The output space is declared in advance, so applications do not need to ask the model to improvise arbitrary prose and then parse it back into a software-safe structure.
03
What Jev can do
TypeSafe positions Jev for software decisions such as classifying or routing a request, scoring urgency or risk, choosing an agent's next tool or subagent, deciding whether an automated workflow should continue or stop, verifying model outputs, and sending uncertain cases to human review.
The important distinction is that Jev does not replace a generative model when an application needs a written response, code or other free-form content. A common architecture is to let Jev make a constrained decision while an LLM handles generation.
04
Jev API
TypeSafe's public API documentation exposes POST /v1/systemone and GET /v1/models. Requests authenticate with a bearer API key. The System One endpoint accepts state plus structured questions and returns typed answers.
Vercel AI Gateway also supports Jev as typesafe-ai/jev. Vercel added AI SDK support on September 16 and expanded access on September 21 with a TypeSafe-compatible client path and a language-agnostic HTTP evaluation API. That makes Jev usable through TypeSafe directly or through Vercel's gateway layer.
05
Jev pricing
TypeSafe currently lists Jev at $0.042 per million input tokens, equivalent to $42 per billion input tokens. Output is listed as free because Jev returns typed values rather than generating long token sequences.
That is the direct vendor price stated at launch. Gateway promotions or third-party routing prices should be treated separately because they can change independently.
06
How fast is Jev?
TypeSafe reports end-to-end response times of roughly 70 to 500 milliseconds and says Jev can be tens to hundreds of times faster and cheaper than frontier LLMs on its System One-shaped workflow evaluations.
Those are vendor-reported benchmarks, not independent measurements. TypeSafe's published comparisons also use workflows designed for the decision-model format, so the numbers should not be generalized to every AI task.
07
Can Jev hallucinate?
TypeSafe uses the phrase 'zero hallucinations' because Jev cannot emit arbitrary strings outside the predefined typed answer structure. That is useful for schema reliability, but it does not mean every decision is correct.
A model can return a perfectly valid typed answer and still choose the wrong option or assign a poor probability. Production systems should calibrate thresholds against labeled examples and keep permissions, validation and irreversible actions in application code.
08
Jev and AI agents
Agent loops create repeated decision points: which tool should run next, whether a task is complete, whether an action is risky, whether to retry, or whether a person should review the result. Jev is explicitly designed for these narrow judgments.
Vercel documents Jev for tool/subagent selection, continue/retry/stop decisions, risk scoring and guardrails. The generative model can still interpret requests and write responses while Jev handles constrained decisions between steps.
09
Availability
Jev launched in early access. TypeSafe's API documentation and console are public, and Vercel AI Gateway now exposes Jev through AI SDK, HTTP and TypeSafe-compatible client paths. Access conditions should still be rechecked before publication because the product moved quickly from launch waitlist language toward broader developer access during its first week.
10
Adoption signal
Vercel reported that Jev became the fastest-adopted model in AI Gateway history during its first 24 hours on the gateway, reaching nearly 13% of paid teams. That is Vercel's first-party platform telemetry, not a measure of the entire AI market, but it is a notable early adoption signal.
11
Limitations
Jev is intentionally not a general-purpose text generator. It is a closed managed model; no open weights or full architecture disclosure were verified. Its calibrated probabilities still need task-specific validation, and its strongest performance claims currently come from TypeSafe's own evaluations.
The model is most compelling where an application already knows the possible decisions and needs a fast probabilistic judgment. If the task requires an open-ended answer, explanation or creative output, an LLM remains the appropriate component.
12
FAQ
Is Jev an LLM?
TypeSafe presents Jev as a different model class called a System One Model. It is optimized for typed probabilistic decisions rather than free-form text generation.
Can Jev generate text?
Not as its intended interface. Jev returns structured decisions rather than prose.
How much does Jev cost?
TypeSafe currently lists $0.042 per million input tokens and free output tokens.
What can Jev be used for?
Routing, classification, scoring, agent-control decisions, guardrails, verification and other constrained software judgments.
Is Jev open source?
No open weights or disclosed full architecture were verified in this run.
Does Jev always make correct decisions?
No. Typed outputs prevent malformed free-form responses, not decision errors. Probabilities and confidence should be calibrated against the application's own data.
Sources
Primary and supporting sources
Use the linked documentation to check current pricing, access and product behaviour. The dates above describe this article; they do not guarantee that every vendor claim remains current.