Startup TypeSafe AI has unveiled its Jev artificial intelligence model – which is designed not to communicate with people, but with other machines. It works much more efficiently and at a much lower cost than traditional large language models.
Photo credit: Numan Ali / unsplash.com
Unlike models that provide answers in natural language, Jev produces probabilistic decisions that can be used by other programs or artificial intelligence systems. This approach is useful in scenarios where interaction with the AI must be limited to a predefined set of response options. For example, a model can play Doom if it is given structured data describing the state of a game character. In practice, however, it will be used to triage support tickets and optimize other business processes.
Using Jev starts by defining the state value – this can be a simple text string or a JSON object; then choose the type of request: choice, numeric, or probability from 0 to 1. For a human such answers are useless, but for software they seem to be the most suitable option. Jev is a System One model based on the reinforcement learning calibration decision (RLCD) architecture.
Image source: Heji/unsplash.com
The most attractive aspect of the model is its efficiency: Where traditional systems can take up to several minutes to prepare a response, Jev takes 70 to 500 milliseconds to respond – a 40 to 200 times improvement in efficiency. In the demonstration video, Jev generated a response in 0.114 seconds, while the OpenAI GPT-5.6 Terra model took 8.566 seconds. Working with Jev, the fee is $0.042 for 1 million input tokens, with weekends free; the same GPT-5.6 Terra is expected to be $2 for 1 million input tokens and $12 for 1 million output tokens. Compared with the flagship Fable 5.1, the cost of the new JEV is 238 times lower.
TypeSafe claims that Jev does not hallucinate, that is, it does not intentionally give wrong answers; but using probabilistic structured answers still does not rule out the possibility of errors. “Contact hallucination systems may be inconvenient for agents, but it is a critical issue for systems with guaranteed latency or deep dependency chains. Existing models, no matter how smart they are, are still prone to hallucinations and typos.””, explained the developer.
Jev is suitable for AI-based automation systems, real-time rapid decision-making applications, classifying large data sets, validating AI inputs, and connecting other AI models.
If you find an error, select it with your mouse and press CTRL+ENTER.
