Gemini tops Grok in Hanoi code duel

This is less a vibe-check than a brutal test of whether an LLM can turn reasoning into working competitive code under protocol, time, and search-pressure constraints. Gemini won because it actually modeled the game; most of the field lost on plumbing bugs or shallow heuristics before strategy even mattered.

• –Gemini was the only bot in the writeup to run real game-tree search with minimax, alpha-beta pruning, iterative deepening, and cycle detection, which is exactly the kind of structure these adversarial puzzles reward.
• –Grok’s one-ply lookahead tied Kimi statistically and looked respectable on paper, but it could not convert that into hero wins against responsive opponents, which exposes the ceiling of shallow search.
• –ChatGPT, GLM, and Nemotron mostly turned the contest into a protocol reliability benchmark by disconnecting, crashing, or never sending moves, a reminder that generated code still fails at the unglamorous parts first.
• –The broader llmcomp standings show big challenge-to-challenge variance, so the takeaway is not “Gemini best at coding” so much as “Gemini best on this exact search-heavy, adversarial coding task.”
• –For developers, the interesting signal is that explicit search and domain-aware evaluation still beat generic “smart” code generation when the problem has tight move budgets and an actively hostile opponent.