Opper's jevman pits six AI models at Pac-Man, incl. GPT-6 Luna Decisions

Jev, made by the US company TypeSafe AI, belongs to a different family from the large language models behind chatbots. Instead of writing prose, a probabilistic decision model takes supplied information and returns a probability for each of a preset list of options — for example, whether a customer inquiry is a return, a technical problem or something else. Because no explanatory text has to be generated, this style of model suits situations where a call has to be made fast.

The benchmark's rules reflect that emphasis. A game ends when all three lives are lost or 5 minutes pass, and the AI receives the maze layout, dot placement and ghost positions as JSON before choosing among up, down, left and right. When a model fails to answer inside its 2-second window, a simple fallback algorithm picks the direction instead, and those substitutions are counted — so a model's score reflects responsiveness as much as judgment.

The games are played against ghosts that follow the same rules as in the traditional Pac-Man, and a model that takes a long time to think its way to a correct answer may not finish near the top. Opper also set the benchmark up so several models, including Kev, Laya, Clef and GPT-6 Luna Decisions, can be compared inside the same game environment, and the official site lets visitors watch AI play or steer Pac-Man themselves against a model.