Jev: TypeSafe’s AI model turns questions into decisions software can use

Published 17 September 2026. Pricing and availability checked on this date.

An incoming customer message might need a reply, but first someone must decide where it belongs, how urgent it is and whether it needs specialist attention. These small judgments keep businesses moving.

Jev, announced by TypeSafe AI on 15 September 2026, is designed for that layer of work. The company, co-founded by former OpenAI researcher Diogo Almeida, presents it as its first “System One” model. TypeSafe’s announcement

What “System One” means

The name borrows from the distinction Daniel Kahneman popularised between fast, intuitive System 1 thinking and slower, deliberate System 2 reasoning. Here, it describes focused judgments made within software, rather than a claim that Jev reproduces human cognition. System One documentation

Jev’s name references economist William Stanley Jevons: TypeSafe’s bet is that cheaper intelligence will encourage more applications, just as greater efficiency can increase consumption of a resource. The naming explanation

An answer your application can act on

Generative models can produce structured outputs too. Jev’s distinction is that it is designed around predefined decisions rather than generating text.

An application supplies context, called its state, and typed questions. Choice selects an allowed option. Score evaluates an ordered rubric. Noul returns a value between zero and one representing the estimated probability that a statement is true. Questions can share one request and are evaluated independently against the same state. API introduction

Instead of asking one broad question such as “What should we do with this customer?”, developers can separate the judgments and combine them using explicit business rules.

A simple example

Illustrative only: this is a hypothetical, simplified response, not a test result.

Input: “Our payment integration stopped working this morning. Nobody can complete an order. Please help.”

The application asks:

• Department: billing, technical or sales?
• Impact: cosmetic issue (0), disruption with a workaround (1), or service blocked (2)?
• Urgency: does the message describe an urgent problem?

A partial structured response could look like this:

{
  "department": {
    "choice": "technical",
    "probabilities": {
      "billing": 0.10,
      "technical": 0.85,
      "sales": 0.05
    }
  },
  "impact": {
    "score": 1.75,
    "probabilities": {
      "0": 0.05,
      "1": 0.15,
      "2": 0.80
    }
  },
  "urgent": {
    "noul": 0.96
  }
}

The score is a weighted position on the defined scale. It is not a measurement of financial damage. The application could prioritise technical review while leaving any refund decision to separate rules. The example follows the documented request-and-response approach and Score semantics.

Where it could fit

Five plausible applications stand out:

• Support routing: select a queue and flag urgency.
• Lead qualification: assess fit and buying intent before prioritising follow-up.
• Content screening: flag suspected policy violations for review.
• Retrieval filtering: score candidate documents before passing context to a generative model.
• Agent orchestration: choose a tool, specialist agent or escalation path.

These are potential designs, consistent with TypeSafe’s use-case documentation, not proof that Jev performs reliably in every setting. A lead score, for example, should be checked against actual sales outcomes.

Confidence is not accuracy

For Choice and Score, TypeSafe derives confidence from the concentration of the returned probability distribution. Noul has no separate confidence field.

A confidence value of 0.95 does not automatically mean 95% accuracy. It summarises the model’s distribution; it does not establish correctness against real outcomes. Confidence documentation

TypeSafe says its training method, Reinforcement Learning for Calibrated Decisions, targets useful probabilities. Calibration concerns groups of predictions: outcomes assigned an 80% probability should occur approximately 80% of the time. That relationship must be tested on representative data. Training and calibration primer

Likewise, a guaranteed output structure cannot guarantee the right decision. Choosing an allowed department incorrectly still sends the ticket to the wrong team.

Pricing, speed and the evidence

As checked on 17 September 2026, TypeSafe lists $0.042 per million input tokens, with output tokens free. Its documentation also warns that rate limits can change during expansion. Current model pricing

TypeSafe reports 70–500 milliseconds end-to-end latency, with published evaluations generally run from West Coast laptops near its service. These are vendor measurements, not a universal response-time guarantee. Performance announcement

Its workflow benchmarks use four tasks and fixed application logic. Reference answers average predictions from GPT-6 Astra and Claude Fable 5.1 at high reasoning effort; other models use provider defaults. Results therefore measure agreement with model-generated references, not independently established ground truth. They cannot establish general business accuracy. Evaluation methodology

Outside testing offers a narrower signal. Every reported a median 0.35 seconds per passage for Jev versus 8.83 seconds for Fable 5.1 at high effort on 12 synthetic passages. Jev caught six of seven intended defects; Fable caught seven. This small experiment illustrates a trade-off, not universal superiority. Every’s reported test

Access and the practical verdict

TypeSafe’s direct service still advertises a waitlist, following its early-access launch. Direct API use requires an account-issued key. Quick start

However, Vercel announced Jev availability through AI Gateway on 16 September. Access is therefore channel-dependent; describing Jev as exclusively waitlisted would miss this development.

Jev currently accepts text, not images, audio or video, and does not write explanations. Performance across languages, ambiguous cases and changing data remains something adopters must evaluate. Current capabilities

Its strongest prospect is complementary: a generative model writes or reasons; Jev supplies bounded judgments; application code controls actions. If the economics and quality hold on real workloads, more software could afford frequent checks. Reliability will still depend on the entire system.