← Return to Benchmark Arena
game-micro-tower-defense-v0 3 / 3 Shipped Tier 2 (Medium / Workhorse)

Micro Tower Defense Game Benchmark

Autonomous Canvas Game generation test across Claude Sonnet 5, Gemini 3.7 Flash, and GPT-5.6 Luna. Evaluated on 60fps game loop, real-time Web Audio procedural synthesis, waypoint pathfinding, tower upgrade economics, and particle effects.

3 Stacks Tier 2 Evaluated
200s – 952s Wall-Clock Span
0 Human Interventions
Sep 2, 2026 Run Date
Real Automated Recording

📊 Benchmark Telemetry & Results

1 Unedited Prompt · Zero Human Follow-up · Measured to Ship
Model Stack Harness Wall Time Speed (TPS) Tokens Est. Cost Build Size Verdict
Gemini 3.7 Flash ⚡
gemini-3.7-flash (medium)
Antigravity CLI 200.0s 339 tok/s 227.1K $0.042 55 KB SHIP ⚡ (Fastest)
GPT-5.6 Luna 🌙
gpt-5.6-luna (medium)
Codex CLI 208.5s 105 tok/s ~38.0K $0.260 20 KB (Compact) SHIP
Claude Sonnet 5 🟠
claude-sonnet-5 (medium)
Claude Code 951.9s 65 tok/s ~82.0K $0.990 48 KB SHIP (Self-Healed)

💰 Token Economics & Cost Breakdown

Official Provider Pricing
Model Stack Input Tokens Output Tokens Thinking Tokens Cached Tokens API Rate (In / Out per 1M) Total Cost
Gemini 3.7 Flash 162,053 65,049 — — $0.10 / $0.40 $0.042
GPT-5.6 Luna ~16,000 ~22,000 — — $2.50 / $10.00 $0.260
Claude Sonnet 5 ~20,000 ~62,000 — — $3.00 / $15.00 $0.990

⚡ Code Quality & Runtime Performance

Automated DevTools Inspection
Model Stack Lines of Code (LOC) Build Footprint Runtime FPS Console Errors Self-Healing Rounds Autonomous Interventions
Gemini 3.7 Flash 1,128 lines 55 KB 60 FPS 0 errors 0 (One-Shot Pass) 0
GPT-5.6 Luna 512 lines (Compact) 20 KB 60 FPS 0 errors 0 (One-Shot Pass) 0
Claude Sonnet 5 1,240 lines 48 KB 60 FPS 0 errors 2 (DevTools Self-Fix) 0
🎮 https://local.app/game-micro-tower-defense-v0/apps/claude.html
Play Fullscreen ↗

📜 The Frozen Benchmark Prompt

Verbatim specification sent to each Tier 2 agent CLI during the autonomous run.

Build a complete, polished, single-file HTML5 Canvas Tower Defense game called `index.html`.

Requirements:

1. Single File Architecture:
   - One standalone `index.html` file only.
   - Zero external dependencies: no external scripts, CDNs, fonts, images, or audio assets.
   - All CSS, JavaScript, and Canvas rendering logic must be inline in the single file.

2. Game Map & Pathing:
   - An HTML5 Canvas rendering a clear winding path from an enemy spawn portal to the player base.
   - Non-path grid tiles where the player can place defensive towers.
   - Enemies follow waypoints accurately along the path curve without drifting off track.

3. Towers & Arsenal (3 Distinct Types):
   - Gun Turret (Cost: 100g): Rapid single-target projectile attack against the nearest or first enemy in range.
   - Cannon (Cost: 175g): Slower firing rate, launches mortar shells that explode on impact dealing Area of Effect (AoE) splash damage.
   - Frost Laser (Cost: 250g): Emits a slowing beam or frost pulse that reduces target movement speed by 40% for a duration.

4. Tower Management & Upgrades:
   - Click/tap on an empty valid tile to open a build menu showing tower choices with gold costs and stats.
   - Click/tap on an existing tower to inspect its range circle, upgrade its damage and range, or sell it for 70% refund.

5. Wave System & Enemy Variety:
   - Progressive waves with countdown timer or "Start Wave" button.
   - At least 3 enemy archetypes:
     - Scout / Runner: Low HP, high movement speed.
     - Grunt: Balanced HP and speed.
     - Behemoth / Tank: High HP, slow movement, rewarding bonus gold.
   - Dynamic HP scaling per wave.

6. Audio & Visual Effects:
   - Procedural sound effects synthesized entirely in real-time via the browser's Web Audio API (shooting, impact/explosion, upgrade, game over). Include a master Mute sound toggle button.
   - Visual juice: projectile animations, muzzle flashes, floating damage numbers, and particle explosions on enemy death.

7. Economy, HUD & State Machine:
   - Top HUD displaying: Player Health / Lives (starts at 20), Current Gold (starts at 450), Current Wave, and Score.
   - Persistent High Score and Highest Wave recorded in `localStorage`.
   - Clean state management: Start / Title screen -> Wave Active -> Victory / Game Over overlay -> Restart Game.
   - Clean, modern, responsive UI overlaying the Canvas with responsive touch & mouse support.

Do not ask any clarifying questions. Build the complete, working, bug-free game file, verify it works in browser, and stop.

🔬 Key Discoveries & Tier 2 Observations

1. Autonomous Self-Verification & Self-Healing Claude Code

Claude Code took longer (951s) because it connected to headless Chrome via Chrome DevTools Protocol, ran automated playthroughs, detected 2 real game-breaking bugs (wave counter stall & restart button state), and fixed them before stopping.

2. Ultra-Fast One-Shot Generation (200s) Gemini 3.7 Flash

Antigravity CLI (Gemini 3.7 Flash) delivered a 55KB rich cyberpunk Tower Defense game with Web Audio synth, floating damage numbers, and 1x/2x/4x speed toggle in only 200.0s.

</> Embed 1Prompt1App Widget

Copy this responsive HTML iframe snippet into your Substack, Blog, or Website: