GPT vs Claude in a bomberman-style 1v1 game

A developer has created an innovative open-source benchmark that pits AI agents against each other in real-time Bomberman-style duels. Unlike traditional static Q&A benchmarks like ARC-AGI 3, this interactive environment tests strategic reasoning by forcing models to make tradeoffs between speed and quality. The game deliberately avoids visual inputs—which remain slow and inaccurate for current models—and instead uses a harness that translates game state into structured text, with the engine rendering responses as fluid animations.

The project represents a shift toward more dynamic AI evaluation methods that reveal practical capabilities and limitations. By having models like OpenAI's GPT-4 and Anthropic's Claude 3 compete directly, developers can observe how different architectures handle real-time strategic decisions. The environment creates genuine tension: smaller models can make more moves but with less strategic depth, while larger models move slower but with smarter reasoning. This approach makes AI capabilities more intuitive to understand through observable behavior rather than abstract scores.

The benchmark is designed to be both informative and engaging, addressing the developer's goal that "benchmarks don't need to be dry bread." The Bomberman format creates clear win conditions where agents must destroy bricks and bomb opponents, providing measurable outcomes. As AI agents become more capable of taking actions in real-world applications, such interactive testing environments will become increasingly valuable for evaluating practical deployment readiness beyond theoretical knowledge.