Best AI model for…
Best AI model for Arcade
Arcade — classic arcade-style game (pick: tetris, breakout, snake).
🥇
Best model for Arcade
GPT-5.6 Sol
"Gorgeous, fully-rendered neon breakout with rainbow brick grid, glowing paddle/ball, retro perspective grid floor, and clean HUD/controls/pause overlay — strong arcade identity backed by solid physics, DPR scaling, particles and audio. Polish and cohesion put it at the top of the field."
See GPT-5.6 Sol's full track record →
The prompt — what I asked each model
Every model on this page got the same fixed prompt inside the Agent Operating System: Arcade — classic arcade-style game (pick: tetris, breakout, snake).
Single HTML file out. No iteration. No "best of N." No examples in the system prompt. Whatever each model produced on the first run is what's on this page. 24 models have attempted it so far — GPT-5.6 Sol, Hermes MoA, Fugu Mini, Fusion, Opus 4.8, DeepSeek V4 Pro, DeepSeek V4 Flash, Claude Fable 5, Fugu Ultra, Fugu Ultra 1.1, Gemini 3.6 Flash, GLM-5.2, Grok, Inkling, Kimi K2.7, Kimi K2.7 · Fast, Kimi K2.7 · No-Think, Kimi K2.7 · Quality, Kimi K3, MiniMax M3, Claude Opus 5, Qwen 3.8, Qwen 3.7, Claude Sonnet 5.
What counts as winning here
This is the game category on Goldie Bench. The question isn't "did the model write code that compiles" — the question is "did the model ship a thing you'd actually use." For Arcade that means three things, in order:
- Does it run? Drop the .html file in a browser. If it opens to a broken page, it scores zero on the first axis.
- Did it hit the brief? The prompt asks for a specific thing. A model that ships a different thing — however polished — gets docked on the brief axis.
- Does it look good? Visual polish, motion, interactivity, attention to detail. This is where the difference between gold and silver usually lives.
Final score is my honest 0–10 across all three axes, averaged. Across the 19 models I've scored on this task so far, the average score is 7.85/10.
Every model's attempt — ranked by my 0–10 score
Models ranked by medal (highest score = 🥇 gold, second = 🥈 silver, third = 🥉 bronze). Click any tile to play that model's actual one-shot HTML.
▶ LIVEGPT-5.6 Sol 🥇
OpenAI
Gorgeous, fully-rendered neon breakout with rainbow brick grid, glowing paddle/ball, retro perspective grid floor, and clean HUD/controls/pause overlay — strong arcade identity backed by solid physics, DPR scaling, particles and audio. Polish and cohesion put it at the top of the field.
▶ LIVEHermes MoA 🥇
Hermes · Mixture of Agents
A polished Neon Breakout with HP bricks, multi-type capsule power-ups (WIDE/SLOW/LIFE), level progression with speed-up, screen shake, flash, lighter-blend particles, starfield, perspective grid, best-score persistence, and full mouse/touch/keyboard control with pause/restart — this edges past Opus 4.8 and Fusion (both 8.5) by combining their juice with deeper systems (HP tiers, power-up variety, level waves) in clean single-file code.
▶ LIVEFugu Mini 🥉
Sakana AI
55KB Breakout-style — paddle, ball, brick wall, particles, score HUD. Smoke-test PASS (large pixel diff after input).
▶ LIVEFusion 🥉
OpenRouter
Neon Breakout with combo system, lives-dot indicators, floating-card overlay animation, full pill HUD. Custom cursor, mobile pointer support, particle juice. Tied with Opus on game-feel.
▶ LIVEOpus 4.8 🥉
Anthropic
All three shipped a genuinely juicy game. Opus's breakout had the most game-feel — particle bursts and a live combo. Kimi's breakout was clean and solid. GLM went its own way with fullscreen neon asteroids. The closest of the practical five.
▶ LIVEDeepSeek V4 Pro
DeepSeek
▶ LIVEDeepSeek V4 Flash
DeepSeek
▶ LIVEClaude Fable 5
Anthropic
Clean, polished neon breakout with glowing rainbow bricks, gradient paddle, twinkling starfield, particle effects, autopilot demo, multi-input controls and level progression — strong and shippable. Slightly generic layout and the ball missed rendering its shadow badge in this frame keep it just under the top tier.
▶ LIVEFugu Ultra
Sakana AI
Ultra v2 — Breakout-style. Smoke-test MAYBE (0.4% diff) — runs but generic click+WASD didn't trigger strong motion; paddle may need mouse-move.
▶ LIVEFugu Ultra 1.1
Sakana AI
The brief asked for a classic arcade game (tetris/breakout/snake); this is a 3D flying-shooter titled 'Breakout Hunter' with drones and radar — polished HUD and clean low-poly visuals, but it completely misses the arcade brief and there's no actual breakout/paddle/brick gameplay visible despite the label. Strong presentation, wrong genre.
▶ LIVEGemini 3.6 Flash
Google
Strong polished HUD, clean 3D player craft with visible enemy and combat systems (projectiles, HP damage taken at 88), but it ignored the brief entirely — this is a twin-stick shooter, not a classic arcade pick (tetris/breakout/snake), and the scene reads sparse with only one enemy visible.
▶ LIVEGLM-5.2
Zhipu / Z.ai
All three shipped a genuinely juicy game. Opus's breakout had the most game-feel — particle bursts and a live combo. Kimi's breakout was clean and solid. GLM went its own way with fullscreen neon asteroids. The closest of the practical five.
▶ LIVEGrok
xAI
A crisp neon Asteroids — glowing ship, lives, score, drifting rocks, real sound. Shippable straight out of one prompt.
▶ LIVEInkling
Thinking Machines
A polished 3D Breakout in Three.js with a gorgeous gradient title, glowing rainbow brick wall, paddle/ball follow, trail dots and live score badge — clearly renders and is on-brief. Held back from top spot by the loose 2D collision math on a 3D perspective view (paddle bounce/wall bounds are hardcoded and can feel off) and lack of lives/win state, but visually it's a strong, shippable entry.
▶ LIVEKimi K2.7
Moonshot AI
All three shipped a genuinely juicy game. Opus's breakout had the most game-feel — particle bursts and a live combo. Kimi's breakout was clean and solid. GLM went its own way with fullscreen neon asteroids. The closest of the practical five.
▶ LIVEKimi K2.7 · Fast
Moonshot AI
▶ LIVEKimi K2.7 · No-Think
Moonshot AI
▶ LIVEKimi K2.7 · Quality
Moonshot AI
▶ LIVEKimi K3
Moonshot AI
Gorgeous neon breakout that rendered cleanly with rainbow bricks, glowing paddle, ball trail, starfield and CRT scanlines, plus a full feature set (powerups, multiball, levels, audio, auto-demo). Very close to the field's best; loses a hair for being a well-trodden brief rather than a novel take.
▶ LIVEMiniMax M3
MiniMax
Neon Breakout — paddle, ball, brick wall, particle trails, score HUD.
▶ LIVEClaude Opus 5
Anthropic
Visually impressive 3D neon arena with polished HUD, minimap, health/boost meters, combat feedback ('HULL HIT -9'), and hostiles on radar — but the brief explicitly asked for a CLASSIC arcade game (tetris/breakout/snake), and this is a 3D twin-stick shooter that ignores the pick entirely. Strong execution but off-brief.
▶ LIVEQwen 3.8
Alibaba
Ambitious 3D reinterpretation of Snake with genuine combat (visible enemies, threats counter, hull/boost systems) and a highly polished neon HUD; strong presentation but it drifts from the classic arcade brief into a bespoke shooter-serpent hybrid, and the scene reads a bit cluttered rather than crisp.
▶ LIVEQwen 3.7
Alibaba
The closest test. All three shipped a real, juicy game. Opus's breakout had the most game-feel (particle bursts + live combo). Qwen's neon breakout is clean and vibrant. GLM went its own way with fullscreen asteroids. Genuinely hard to separate.
▶ LIVEClaude Sonnet 5
Anthropic
Clean, polished neon Breakout that renders perfectly with glowing bricks, gradient paddle, particle bursts, and a live in-progress state (score 10, a brick already cleared). Solid physics with side-aware collisions and level scaling, but it's a well-executed take on the most common pick rather than something that clearly tops the best-in-field.
How I tested this — the methodology in 60 seconds
Every comparison on Goldie Bench follows the same recipe:
- I pick a creative coding prompt that a frontier model should be able to ship in one shot.
- I dispatch the exact same prompt to each model from the kanban inside the Agent Operating System.
- I save whatever .html file the model produced on the first run. No iteration. No coaching.
- I score each result 0–10 on my three axes (runs / hits the brief / looks good).
- I publish the scores publicly in the source comparison guides on agentos.guide — and that's what feeds this page.
See the methodology page for full data provenance, including which source guides each cell's score came from.
Related tasks and comparisons
More tasks in the Game category · All attempts on Arcade · Back to the leaderboard
The same stack Julian uses
Run this stack yourself.
Every demo on this bench was built inside the Agent Operating System — one prompt, one shot, single HTML file out. The Agent OS, the prompts, the templates, the weekly walkthroughs and 4,000+ founders shipping with it every day all live inside the AI Profit Boardroom.
4,000+founders
258documented wins
38countries
$59/momonthly