Jev AI: Turn Questions into Structured Decisions

Jev AI is TypeSafe AI’s first System One model. Instead of generating free-form text like ChatGPT, it takes unstructured state (text or JSON) plus typed questions and returns structured decisions — choices, scores, or yes/no probabilities — with calibrated confidence. It is designed for software to act on directly.

Traditional LLMs generate text token-by-token for humans to read. Jev AI never writes prose, code, or explanations. It performs parallel structured decisions in one pass, delivering type-safe outputs with no hallucinations of format, much lower latency (typically 70–500 ms), and far lower cost for classification, routing, and scoring tasks.

No. While it is transformer-based and understands language, Jev AI is not a generative language model. It does not produce text at all — only pre-defined typed answers with probabilities. TypeSafe calls this category a “System One Model.”

These are the three core question types: - Choice: Selects one option from a list you define (up to 255 options) and returns the full probability distribution. - Noul: Answers a yes/no question with the probability that the answer is yes. - Score: Rates the input on an ordered scale/rubric you set and returns probabilities plus a weighted score. You can mix multiple questions of any type in a single call.

TypeSafe reports end-to-end latency of 70–500 milliseconds — claimed to be 40–200× faster than frontier LLMs on equivalent structured decision tasks. Real-world tests often show median times around 250–400 ms for typical requests.

Official pricing is $0.042 per million input tokens. Output tokens are free because the response is just a few numbers. A typical short request often costs around $0.00002. Many independent playgrounds (including competitors) offer limited free trials.

Jev AI cannot invent labels or return malformed output because every possible answer is defined in advance by you. However, it can still select the wrong option. That is why the calibrated confidence scores matter — use them to route low-confidence cases to a human or fallback system.

Common high-value uses include: - Support ticket routing and triage - Agent tool selection and loop control - Risk scoring, spam/safety filters, and guardrails - Lead quality scoring - Verifying or judging outputs from other models - Real-time decisions in games, simulations, or automation It excels at high-volume, bounded judgments inside software.

Jev AI was built by TypeSafe AI, a San Francisco startup founded by Diogo Almeida (former OpenAI researcher who co-invented RLHF and worked on InstructGPT). The company raised a $40M seed round led by DCVC and emerged from stealth in September 2026.

Yes — independent sites such as jev-ai.net offer a free playground with no credit card required (subject to rate limits). You can experiment with Choice, Noul, and Score examples immediately. For production API access, use TypeSafe’s official channel or third-party gateways (OpenRouter, Vercel AI Gateway, Cloudflare).

No. Jev AI is deliberately non-generative. It only returns the typed decision and probabilities you requested. Pair it with a traditional LLM when you need explanations, writing, or complex multi-step reasoning.

No, the model weights are not open source (though some community reproductions exist). TypeSafe states that customer requests and responses are not used for training or fine-tuning. You customize behavior through the state, question wording, and criteria you provide.