Which LLM Builds the Best Web Gomoku Game? A Comparative Test of the Latest Models

Overall Analysis

Using the same prompt “Implement a web‑based Gomoku game”, the author fed six recent LLMs (deepseek‑v4‑pro, Opus 5, gpt‑5.6‑luna, gpt‑5.6‑sol, glm‑5.2, MiniMax‑M3) through the pigo CLI with a medium reasoning setting. All models produced a playable Gomoku game, but they diverged clearly on four criteria: whether an AI opponent is included, how the code is run, visual design style, and overall engineering rigor.

Comparison Overview

deepseek‑v4‑pro : single‑file index.html, double‑click to run, no AI (pure two‑player), dark‑blue theme with orange title, clear move log. Very stable and minimal.

Opus 5 : modular project (multiple JS files, server.mjs, package.json), requires npm start and a local server, includes alpha‑beta AI with difficulty selection, accessibility (aria‑live), high‑DPI rendering, and unit tests. Most engineering‑heavy.

gpt‑5.6‑luna : three files, double‑click to run, no AI, magazine‑style visual layout (light‑brown background, large Chinese slogan), visually striking but functionally limited to two‑player.

gpt‑5.6‑sol : three files plus artifacts, double‑click to run, includes AI (challenge‑computer mode), clean light theme, rule description, timer, keyboard controls, and accessibility support.

glm‑5.2 : single‑file, double‑click to run, no AI (placeholder), dark theme with timing cards, performs self‑validation after generation (JS syntax check, win‑condition tests, headless Chrome screenshot, DOM checks) before presenting the result.

MiniMax‑M3 : three‑file layout with a polished coffee‑brown UI, implements heuristic AI in code, but human‑vs‑AI mode never makes a move, leaving the game locked; only the two‑player mode works.

Per‑Model Details

deepseek‑v4‑pro

The model followed a “single‑file delivery” plan, listing five steps (board, turn‑taking, win detection, undo/restart/score, UI polishing) and packed everything into a ~22 KB index.html. The interface shows a 15×15 board with coordinates, a side panel for current player, score, undo, and a move log. Features include undo (Ctrl+Z), last‑move highlight, and responsive layout. The author notes it is the most stable and quickest to run, but lacks an AI opponent.

Opus 5

Opus 5 splits the code into gomoku.js, ai.js, styles.css, main.js, index.html, plus package.json and a zero‑dependency server.mjs. Because it uses ES modules, the file cannot be opened directly; the author runs npm start and visits localhost:8000. The dark theme includes a subtitle “15×15 board · two‑player or human‑vs‑AI · no forbidden moves”. It offers the most complete UI: mode switch, AI difficulty, move history, scoring, accessibility (aria‑live), high‑DPI rendering, and keyboard controls. It also provides npm test unit tests, making its engineering quality the highest among the samples.

gpt‑5.6‑luna

The visual design mimics a magazine spread: light‑beige background, large Chinese slogan “No regret in placing stones, victory in heart”, English tags “GOMOKU / TRADITIONAL BOARD GAME”. Functionally it offers only local two‑player play (current turn, move log, undo, restart) in three files ( index.html, styles.css, app.js) with no dependencies. The author describes it as the most eye‑catching UI, suitable for demos.

gpt‑5.6‑sol

This model follows a product‑oriented approach. The top bar toggles “Challenge computer / Player vs player” and a sound switch. The left panel shows player info, turn, timer, and score; the right panel displays rule explanations (black first, five‑in‑a‑row wins, no forbidden moves) and hints for keyboard interaction. Delivered as index.html, styles.css, game.js plus an artifacts folder, it includes a working AI, win detection, undo, scoring, accessibility, and keyboard support.

glm‑5.2

After generating a single‑file index.html, glm‑5.2 runs an automatic self‑check: JavaScript syntax passes, win‑condition tests for four directions succeed, a headless Chrome screenshot (720×900, 274 KB) confirms rendering, and key DOM elements are verified. The dark theme shows timing cards for both sides, high‑DPI canvas, coordinates, and a red circle on the last move. It supports two‑player play, undo, restart, keyboard shortcuts, and draw detection, but AI is only a placeholder.

MiniMax‑M3

MiniMax‑M3 presents the most refined UI: coffee‑brown three‑column layout with mode/stone selection, board, and scoreboard. The board uses a wood‑grain texture and animated win effects. The code contains a heuristic AI (evaluating five‑in‑a‑row, live‑four, etc.) but in practice the human‑vs‑AI mode never makes a move: after the black stone is placed, the white AI stays “thinking” indefinitely, locking the game without errors. Only the two‑player mode works.

Conclusions

Quickest to run, minimal effort : deepseek‑v4‑pro and glm‑5.2 (single‑file double‑click).

Human‑vs‑AI needed : Opus 5 (engineered with tests) and gpt‑5.6‑sol. MiniMax‑M3’s AI is broken.

Best visual design : gpt‑5.6‑luna (magazine style) and MiniMax‑M3 (polished three‑column).

Engineering rigor : Opus 5 (modular, unit‑tested) and glm‑5.2 (self‑validation).

An important observation is that, given the same vague requirement, different models make very different assumptions about missing details. Some stop at two‑player play (deepseek, luna, glm), while others proactively add AI, difficulty levels, rule explanations, and even automated tests (Opus 5, sol, MiniMax). The author warns that a model’s listed feature set cannot be trusted without actually running the code.