What is Jev AI? Pricing, use cases and limitations

What is Jev AI?

Jev is TypeSafe AI's model for making structured decisions. You send text or JSON describing a situation, together with questions whose answer types you define. It returns choices, scores or probabilities that your code can use. TypeSafe calls it a System One model. Model introduction.

Think of the decisions that happen before an AI assistant replies: which queue should receive a request, how urgent is it, and is there enough evidence to proceed? Jev is built for judgments like these. It does not itself grant permission to act.

Developer

TypeSafe AI

Launch

September 15, 2026. Launch announcement.

Outputs

Choice, Score and Noul. Question types.

Published price

$0.042 per million input tokens; output tokens free. Pricing source.

Input

Text, including JSON; no direct image, audio or video input for the documented Jev 1.13 model. Input limits.

Our assessment

A candidate for frequent, bounded judgments. This is a documentation-based guide, not our own performance benchmark.

How Jev works: Choice, Score and Noul

Every call pairs a shared state with a set of typed questions. According to TypeSafe, the questions are evaluated independently and in parallel against that state. Interface documentation.

• Choice: select an option from a list, such as the team that should handle a request. The response includes the choice, a distribution over options and a confidence value.
• Score: evaluate against an ordered rubric you define, such as incident severity. The response includes a score, a distribution over rubric levels and confidence.
• Noul: estimate the probability that a statement is true, such as whether a message expresses urgency. The result is a number between zero and one.

These are the three documented primitives. Your application decides how to combine their outputs.

A worked example: routing a support request

Application rules can then choose a queue and escalation path. A generative model can draft the reply. These are illustrative steps, not outputs measured in a Jev test.

Separating those decisions has a practical benefit: changing the escalation policy does not require rewriting the instructions for composing an apology.

Jev vs an LLM with structured output

Language models already support structured output. Anthropic, for example, documents constrained JSON and strict tool inputs. Comparing Jev only with unconstrained chat would miss a real alternative. Structured-output documentation.

The interface descriptions come from TypeSafe and Anthropic; the suggested roles and evaluation criteria are our editorial judgment.

TypeSafe describes a specialized architecture, parallel sampling and Reinforcement Learning for Calibrated Decisions (RLCD). It reports 70-500 ms end-to-end latency. These are vendor descriptions and measurements, not results independently reproduced here. Architecture and timing claims.

The strongest counterargument is simple: a small language model, conventional classifier or deterministic rule may already meet your needs. Jev has to earn its place on your workload.

Jev pricing: what does an evaluation cost?

TypeSafe's published price, checked September 22, 2026, is $0.042 per million input tokens, with free output tokens. Confirm your selected provider's current terms. Model pricing.

For an illustrative workload of 100,000 evaluations using 1,000 total billable input tokens each:

This excludes retrieval, retries, other models, infrastructure and human review. Cheap evaluations can still produce an expensive workflow if they send the wrong cases to a fallback or create work someone must undo.

TypeSafe's large speed and cost multipliers come from its own workflow evaluations. It acknowledges possible evaluation bias and says the gains are toward the high end of expected real-world improvements. Benchmark caveats.

For a deployment decision, measure the total cost of finishing the job correctly.

Where Jev could fit in an AI agent stack

These are candidates to evaluate, not measured recommendations or promises of better accuracy.

• Request routing: interpret intent, then let application rules choose a queue, tool or model. Keep a fallback for mixed or unfamiliar requests.
• Retrieval filtering: score whether a passage addresses a query before adding it to context. Test for lost relevant evidence, not just shorter prompts. See our RAG stack guide.
• Workflow triage: score a case against explicit criteria and send uncertainty to review. Our AI workflow guide covers the surrounding orchestration.
• Agent checks: flag suspicious proposed actions for closer inspection. Preserve hard permissions and approval rules in the coding-agent harness.

Vercel documents Jev on AI Gateway as typesafe-ai/jev, with access through the AI SDK's experimental evaluation API. Its announcement describes routing, risk scoring and output checking as use cases. Access and integration.

Jev's limitations: a valid answer can still be wrong

TypeSafe's zero-hallucination figure reflects guaranteed schema matching. It is not an empirical result showing zero decision errors. The qualification behind the claim.

If the allowed options are A, B and C, the model can return B in exactly the required format while A is correct. The software accepts the response; the judgment is still wrong.

What the confidence score tells you

TypeSafe trains for calibration: across comparable predictions, outcomes assigned a probability of 0.8 should occur about 80% of the time. That is a group-level property, not a guarantee about an individual answer. Calibration explained.

For Choice and Score, the confidence field summarizes the shape of the probability distribution. It is not an independent second opinion about correctness. Noul has no separate confidence field. Confidence documentation.

Choose escalation thresholds from observed performance and the consequences of mistakes. A help-page suggestion and an irreversible action should not inherit the same authority.

Documented failure modes

For Jev 1.13, TypeSafe lists difficulties with counting, numerical precision, date comparisons, indirect reasoning and distracting context. It also warns that adversarial input can steer answers and that text generation is a poor fit. Version-specific limitations.

Our recommendation: keep exact calculations in code, make the relevant context explicit, and enforce permissions outside the model. A probabilistic check can inform a security policy; it cannot substitute for the policy.

How to evaluate Jev before using it

Start with one reversible decision and a labeled set of examples. Keep the same cases and success criteria for Jev and your existing approach. Include the straightforward cases and the awkward ones you cannot afford to miss.

The worksheet below is an original starter test set for the support-routing example. All results are blank. It is designed to reveal failure modes, not establish a statistically reliable production error rate.

Download the Jev evaluation worksheet (CSV)

1. Fix the task. Write allowed outcomes, review rules and the mistakes that would stop deployment.
2. Keep a held-out set. Tune questions on separate examples, then evaluate without rewriting them to fit each answer.
3. Record the whole path. Capture model version, errors, latency, input tokens, fallbacks and human review.
4. Compare useful completions. Count missed urgent cases and unnecessary escalations separately. Report cost per correctly completed case.

A fast model that misses the cases you most needed to catch may produce a worse system. The useful question is which decisions are frequent, bounded and testable enough to delegate.

Frequently asked questions about Jev AI

What is Jev AI?

Jev is TypeSafe AI's model for structured decisions. You supply text or JSON state and typed questions; it returns choices, scores or probabilities that software can use. TypeSafe calls this a System One model. Source.

How much does Jev cost?

TypeSafe lists Jev at $0.042 per million input tokens, with free output tokens, as checked September 22, 2026. Total application costs can also include retrieval, retries, other models, infrastructure and review. Check your provider's current billing terms. Source.

Can Jev write code or replace ChatGPT?

Jev's interface is designed for bounded decisions rather than composing text or code. It can help judge or route work around a generative model. TypeSafe advises using a generative model when the task requires text generation. Source.

Does Jev eliminate hallucinations?

TypeSafe's zero-hallucination claim refers to schema matching, not guaranteed correctness. Jev can choose the wrong option from an allowed list. A valid output format does not prove that the underlying judgment is right. Source.

How can developers access Jev?

TypeSafe provides a hosted API. Vercel also documents access through AI Gateway as typesafe-ai/jev, using the AI SDK's experimental evaluate API. Confirm access, quotas and billing with the provider before integrating it. Source.

Sources and editorial method

We reviewed the following primary sources on September 22, 2026. Product descriptions, prices and performance figures are attributed to their publishers. The workflow design and evaluation method are our editorial analysis.

We have not run a Jev benchmark for this guide. Confidence is high in the account of the cited documentation; expected benefits for a particular application remain unverified. Illustrations are original AI-generated artwork.

• TypeSafe: Introducing System One Models and Jev
• TypeSafe: model introduction
• TypeSafe: Choice, Score and Noul
• TypeSafe: model versions, pricing and inputs
• TypeSafe: confidence and probability
• TypeSafe: training for calibrated decisions
• TypeSafe: Jev 1.13 limitations
• Vercel: Jev on AI Gateway
• Anthropic: structured outputs

Read our editorial methodology and agent framework selection guide. Organic coverage is independent of sponsorship.