What Is Jev? The $0.042 Decision Model

What Is Jev? The $0.042-per-Million-Token Decision Model

Jev is TypeSafe AI’s first public System One Model, an early-access AI model that evaluates typed questions against application state and returns bounded decisions with probabilities instead of generated text. TypeSafe released Jev on September 15, 2026, with a published price of $0.042 per 1 million input tokens and free output. TypeSafe describes response times of 70–500 milliseconds for System One tasks.

Key Takeaways

• Jev is designed for software decisions, not conversation or open-ended text generation.
• Its interface accepts state plus typed questions and returns structured answers, probabilities, and confidence values.
• The three core question types are Choice, Score, and Noul.
• TypeSafe AI says Jev evaluates questions in parallel rather than generating one token after another.
• TypeSafe reports up to 193.6× faster and 444.6× cheaper performance on selected System One workflow comparisons.
• Jev is available through TypeSafe AI’s early-access API, while downloadable weights and a public license are not yet confirmed.

What Is Jev?

Jev is an AI evaluation model from TypeSafe AI. The company calls the category a System One Model, borrowing the idea of fast, intuitive judgment from Daniel Kahneman’s distinction between System 1 and System 2 thinking.

The practical distinction is architectural and operational. A conventional language model receives context and generates a string. Software then parses that string, validates it, and decides what to do. Jev receives application state and a set of questions whose answer space is defined in advance. It returns the answer directly in a typed structure.

Jev is best understood as a probabilistic decision function for software rather than as a chatbot.

A request might ask whether a support ticket is urgent, which team should handle it, or whether an agent’s last action completed its goal. The application defines the questions and allowed outputs. Jev supplies the judgment and probability distribution. Code remains responsible for thresholds, policies, retries, and final actions.

TypeSafe AI introduced Jev after two years of development in stealth under founder Diogo Almeida, who previously worked on instruction-following and reinforcement-learning research at OpenAI. The company describes its training approach as Reinforcement Learning for Calibrated Decisions, or RLCD. RLCD is intended to make confidence values useful as operating signals, so a system can distinguish high-confidence automation from cases that need review.

Jev at a Glance

Specification	Jev
Developer	TypeSafe AI
Type	System One Model for structured decisions
Input	Text, structured objects, or arrays representing application state
Output	Typed decisions, probabilities, and confidence values
Question types	Choice, Score, and Noul
Context window	Not yet confirmed
Response time	70–500 milliseconds, according to TypeSafe AI
Input pricing	$0.042 per 1 million input tokens
Output pricing	Free, according to TypeSafe AI
Availability	Early-access TypeSafe API
Downloadable weights	Not yet confirmed
License	Not yet confirmed
Public release	September 15, 2026

The published input price equals $42 per 1 billion input tokens. That figure excludes application-side costs such as storage, orchestration, data preparation, and any other models used alongside Jev.

How Jev Works / What Makes It Different

The Jev API separates state from questions. State can be a ticket, record, log, document, conversation, or structured application snapshot. Questions describe the judgments the software needs.

The first primitive, Choice, selects one item from a declared set. A ticket can be routed to billing, technical support, or sales. The result includes the selected option, probabilities for the available options, and a confidence value.

The second primitive, Score, rates state against an ordered rubric. A system might score urgency across five levels or assess the quality of an answer against several ordered descriptions.

The third primitive, Noul, answers a yes-or-no question as a probability from 0 to 1. A Noul question can ask whether a message contains a prompt injection, whether a transaction looks suspicious, or whether an agent goal has been completed.

All three question types can be combined in one request. TypeSafe says the questions are evaluated in parallel and in isolation against the same state. This means a request can ask several independent questions without requiring a separate conversational turn for each one.

The main speed advantage comes from avoiding ordinary token-by-token answer generation. TypeSafe describes Jev as using a parallel sampler and a model stack built for structured outputs. The company does not publish enough technical detail to independently reconstruct the full architecture, parameter count, or hardware implementation.

RLCD is the second major differentiator. Standard language-model confidence is often inferred from token probabilities or prompted self-reporting. Jev’s design makes probabilities part of the decision interface. Developers can then implement Confidence-Gated Routing, such as automatically approving results above 0.9, sending borderline cases to another model, and escalating low-confidence cases to a human.

Bounded output also reduces schema failures. If a question allows only approved options, Jev cannot emit an invalid option or malformed JSON for that field. That does not make every judgment correct. A model can still select the wrong valid answer or attach an overconfident probability to it.

What You Can Do With Jev

Jev fits workflows with repeated, narrow judgments and a known output space.

• Agent routing: classify a request and select the appropriate model, tool, sub-agent, or reasoning depth.
• Safety checks: decide whether a command, tool call, prompt, or retrieved document should be allowed, blocked, or reviewed.
• Verification: check whether an agent completed its goal after each turn. Community testing describes this as a way to run more frequent verifiers at lower cost.
• Content and ticket classification: route customer messages, score urgency, detect escalation signals, and identify likely departments.
• Browser automation: the community project Jev Ultrafast uses Jev to choose browser operations and targets from an indexed action list. A smaller language model supplies text only when a browser field requires arbitrary typing.
• Code and agent evaluation: the model can act as an evaluator for risk, severity, evidence support, or completion status. Community projects also connect it to MCP clients and code-review workflows.
• Runtime UI composition: a demonstration with predefined UI components used Jev to choose component combinations and ordering, while ordinary code performed the rendering.
• Scoring pipelines: TypeSafe documentation shows composite scoring, where separate dimensions are evaluated and weighted in application code.

A browser demo reported a complete Google Flights flow in about 7.1 seconds, but that result belongs to the community application, not to Jev alone. The browser, text model, network, page load, and orchestration all affect the total.

How Jev Compares

Jev competes with general-purpose language models only on a narrow class of tasks. It is not a replacement for a model that must explain an answer, write code, plan several steps, or generate user-facing language.

Dimension	Jev	Conventional frontier LLM
Output	Typed bounded values and probabilities	Generated text or structured text
Sampling	Parallel evaluation, according to TypeSafe AI	Sequential token generation
Published response range	70–500 milliseconds	3–329 seconds in TypeSafe’s comparison
Input price	$0.042 per 1 million tokens	$0.20–$10 per 1 million tokens in TypeSafe’s comparison
Best fit	Classification, routing, scoring, verification	Reasoning, writing, coding, planning
Main limitation	Cannot provide open-ended language	Slower and more expensive for small decisions

TypeSafe reports up to 193.6× faster and 444.6× cheaper results across four selected workflows. Those are vendor-reported comparisons, not an independent standardized leaderboard. A third-party test can produce different results because latency depends on batching, hardware, network distance, prompt length, concurrency, and the exact competing model.

For readers comparing general-purpose systems, the site’s earlier analysis of GPT-6 Astra and comparison of GPT-6 Astra with Claude Fable 5.1 cover a different workload category. Jev belongs beside those models in an architecture, not necessarily instead of them. A system can use Jev for fast routing and verification, then send selected tasks to a general model such as GPT-6 Astra for reasoning or generation.

Jev’s advantage is strongest when a software system needs many bounded judgments, not when it needs fewer but more expressive answers.

Availability: How to Access Jev

Jev is available through TypeSafe AI’s early-access API. TypeSafe’s public API documentation identifies the jev-latest model and the /v1/systemone evaluation interface. Access requires a TypeSafe API key, and early access has involved a waitlist.

TypeSafe’s documentation is the primary source for current request formats, supported fields, limits, and terms. The API accepts state plus a map of typed questions. It returns answers under the question identifiers supplied by the developer.

Jev has also appeared through third-party developer integrations, including an AI gateway and MCP servers for coding agents. Those integrations can simplify experimentation, but they do not change Jev’s underlying availability, pricing, or terms. The vendor’s API remains the reference point for production use.

The model is not listed as downloadable weights, and TypeSafe has not published an open-source license for Jev. Developers who need local inference can examine independent projects, but compatibility with the Jev service does not prove equivalent accuracy or calibration.

What We Don’t Know Yet

Several important engineering facts remain unpublished or difficult to verify independently:

• The model’s parameter count, architecture diagram, training corpus, and hardware requirements are not public.
• The context-window limit is not yet confirmed in the available documentation.
• TypeSafe has published headline speed and cost comparisons, but independent evaluations with reproducible methodology remain limited.
• Calibration quality across domains, languages, adversarial inputs, and distribution shifts needs broader testing.
• Long-term pricing, rate limits, service-level commitments, and general-availability timing are not fully established.
• The license and deployment rights for any future Jev weights are not confirmed.
• Viral demos can hide preprocessing. Browser and game demonstrations often convert raw environments into structured state before Jev sees them.
• TypeSafe’s terms reportedly restrict public benchmark or performance disclosures. That policy is based on community reporting and should be checked against the current terms before publishing measurements.

Independent Jev-like projects are emerging quickly. OpenJev provides an open-source compatible decision server, while SemIf reads option probabilities from open models. These projects are useful for research, but they should not be treated as evidence that the TypeSafe model can be replicated without proprietary training data or infrastructure.

Frequently Asked Questions

What is Jev?

Jev is TypeSafe AI’s first public System One Model, designed to evaluate typed questions against application state and return bounded decisions with probabilities instead of generated text. It is built for software workflows such as classification, routing, scoring, verification, and agent control.

Is Jev an LLM?

Jev is not a conventional text-generating LLM, although it is an AI model that can evaluate natural-language state. It returns typed answers and probability distributions rather than producing a conversational response token by token.

How much does Jev cost?

Jev costs $0.042 per 1 million input tokens, while output tokens are listed as free by TypeSafe AI. The long-term pricing and any access-specific limits should be checked against the latest TypeSafe terms.

Is Jev open source?

Jev is not open source, and TypeSafe AI has not published downloadable Jev weights or an open license. Independent projects such as OpenJev, SemIf, and other Jev-compatible implementations exist, but they are not the TypeSafe model.

What can Jev be used for?

Jev can classify records, select tools or routes, score inputs, verify agent results, assess risk, and make other bounded semantic decisions. It is not designed for chat, long-form writing, code generation, or open-ended explanations.

Jev vs GPT-6 Astra?

Jev is optimized for fast, bounded software decisions, while GPT-6 Astra is a general-purpose chat model in the available comparison set. Jev is the better fit when the answer space is defined in advance; GPT-6 Astra is the better fit when the system must reason, write, explain, or generate code.

Can Jev generate text?

Jev does not generate arbitrary text as part of its System One interface. It returns Choice, Score, and Noul results that software can consume directly.

What to watch next: track TypeSafe’s move from early access to general availability, independent calibration studies, and whether open local decision models approach Jev’s quality on the same tasks.

Try related models on Kie.ai: Claude Haiku 5.5, GPT-6 Astra and DeepSeek-V4.1-Flash.