Jev: A New AI Model From a ChatGPT Creator That Returns Probabilities, Not Text
Published · Updated
Cited: Forbes, TechCrunch, MarkTechPost
TL;DR
TypeSafe AI's Jev model outputs typed decisions with probabilities, achieving up to 100x lower cost and 200x faster inference compared to traditional LLMs for classification tasks.
Why now
As AI adoption surges, the need for cheap, fast, and reliable decision-making models becomes critical. Jev fills a gap by offering a non-hallucinating, confidence-scored alternative to expensive LLMs for routing, safety, and automation.
Agree / conflict
While Jev's speed and cost are impressive, its benchmarks are vendor-run and may not reflect real-world gains. Skeptics note that the model is proprietary and lacks transparency.
Takeaway
Developers can prototype with Jev's API for free and potentially slash inference costs by orders of magnitude for classification tasks.
TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, has released Jev, a transformer-based model that outputs calibrated probabilities rather than text. This design eliminates hallucinations and enables cheap, fast inference—$42 per billion input tokens and response times under 500ms. Early adopters like Vercel and Bryo AI report dramatic cost and speed improvements versus GPT and Gemini.
Jev is particularly suited for software automation tasks such as email triage, command safety, and agent routing. Its three primitive question types (Choice, Score, Noul) allow developers to integrate decision-making without the overhead of a full LLM. However, the model is only available via API under a waitlist, and its architecture remains undisclosed. TypeSafe positions Jev as a 'System One' model, optimized for fast intuition, and plans to release additional modalities.
FAQ
What is Jev?
Jev is a transformer-based AI model from TypeSafe AI that outputs typed decisions with probabilities, not text. It is designed for fast, cheap, and reliable decision-making tasks.
How does Jev compare to large language models?
Jev is up to 200x faster and 100x cheaper than some LLMs for classification tasks, but it is limited to predefined outputs and lacks generative capabilities.
Is Jev open source?
No. Jev is proprietary and available only through a hosted API under a waitlist. Weights are not released.
Can Jev hallucinate?
Jev cannot produce invalid outputs because it returns only user-defined typed decisions. However, the probabilities can be wrong, so accuracy depends on the task.
What are the costs of using Jev?
Input tokens cost $42 per billion (approximately $0.042 per million). Output tokens are free. End-to-end response times range from 70ms to 500ms.
Who is likely to benefit most from Jev?
Developers building automated systems that require cheap, fast classification—such as email routing, command safety, agent guardrails, and data filtering.
Sources
Canonical URL: /trend/2026/jev-a-new-ai-model-from-a-chatgpt-creator-that-returns-probabilities-not-text
Disclaimer
Digests summarize public sources. They are not advice, forecasts, or a complete record of every trend.
This digest was compiled by Yanuki using publicly available data and trending information. The content may summarize or reference third-party sources that have not been independently verified. While we aim to provide timely and accurate insights, the information presented may be incomplete or outdated.
All content is provided for general informational purposes only and does not constitute financial, legal, or professional advice. Yanuki makes no representations or warranties regarding the reliability or completeness of the information.
This digest may include links to external sources for further context. These links are provided for convenience only and do not imply endorsement.
Always do your own research (DYOR) before making any decisions based on the information presented.