A Reddit user this week used Pac-Man to host a tournament among six AI models. The champion was Jev 1.13 with an average score of 2,750; the runner-up was GPT-6 Luna with an average latency of 179 ms. It sounds like geeks amusing themselves, but we noticed something: the contestants weren't the familiar chat-based large models, but a new species called "decision models" (small models specialized in determining the next action) — response times generally between 100 and 400 ms, fast enough to play games in real time.
What This Is
Decision models are a category distinct from "chat models." Chat large models (like GPT, Claude) excel at understanding long-form text and generating content, but responses typically take several seconds to over ten. Decision models sacrifice some general capability in exchange for millisecond-level response speed, specializing in tasks like "deciding what to do next."
The six models in this tournament — Jev 1.13, Kev 4B, Clef, Clef Flash, GPT-6 Luna, and Laya — had average latencies ranging from a fast 104 ms to a slow 398 ms. That means they can be embedded in scenarios requiring immediate feedback: game NPCs, customer service routing, automated trading, risk-control interception, and robot control.
The project team open-sourced the leaderboard and code, running each model 100 times to compute the mean, and encourages users to submit their own fine-tuned models for scoring. This is an atypical use of games as benchmarks, similar to how Go was once used to test algorithmic capability.
Industry View
Supporters believe this points to a clear trend: models no longer have just one path of "big and comprehensive." Small models + strong specialization + low latency are becoming another quadrant of AI productization. Cloudflare's launch of Clef and OpenAI's release of a decisions API are extensions of this line — turning "decision-making" from a side project of large models into an independent product.
But the opposing view is equally valid. Game scores may not map to real business value; 290 ms feels "fast," but it's still slow for high-frequency trading and industrial control. More critically, the decision complexity of Pac-Man and real-world tasks are not in the same league — whether a model that scores 6,000 points can handle contract clause disputes is a different question entirely.
Another view holds that this is "self-amused" innovation from the open-source community in the absence of recognized evaluation standards. Entertaining in the short term, but in the long run, more rigorous benchmarks (such as tool-calling accuracy and long-chain task completion rates) are needed before decision models can truly enter enterprise procurement lists.
Impact on Regular People
For enterprise IT: If decision models deliver stable output, many in-enterprise automation scenarios that currently "call an API and wait three seconds" will see experience upgrades — especially high-frequency small decisions like customer service routing and marketing triggers.
For individual professionals: The AI assistants you use in the future may split into two categories — "write and think" large models handle the thinking, while "fast, accurate, ruthless" decision models handle the doing. They'll divide the labor rather than one model carrying all tasks.
For consumer markets: "AI opponents" or "AI customer service" in games, e-commerce, and customer service apps will become smarter and more immediate, reducing the "AI is playing dumb" phenomenon caused by high latency.