Claude Code vs GPT-5 (Cursor): My 3D Aquarium Test on Launch Day

Today's a big day — Aug 7, 2025 — the day GPT-5 officially dropped.

Naturally, I had to test it. But instead of a quick coding prompt, I decided to run it through the same one-shot challenge I gave Claude Code (Sonnet 4): build a detailed 3D aquarium app.

Here's the twist: both GPT-5 (through Cursor) and Sonner 4 (through Claude Code). But after running the test, I walked away still preferring Claude.

The Test

I handed each AI the exact same prompt — no back-and-forth, no retries. Just one go at designing and implementing:

• A glass-walled aquarium with sand, lights, and water effects
• Five different fish species, each with unique behaviors
• Seaweed, bubbles, and subtle water currents
• Interactive camera controls
• A performance-friendly, production-ready setup

3D Aquarium Implementation Plan
1. Dependencies Setup
   - Install Three.js (three) and its TypeScript types (@types/three)
   - Install React Three Fiber (@react-three/fiber) for React integration
   - Install React Three Drei (@react-three/drei) for additional 3D components and controls

2. Core 3D Scene Structure
   - Create components/Aquarium3D.tsx as the main aquarium component
   - Set up Canvas with camera, lighting, and controls
   - Implement OrbitControls for drag-to-rotate, zoom, and pan interactions

3. Aquarium Environment
   - Tank: Create transparent glass walls using BoxGeometry with glass material
   - Floor: Add sandy textured floor with noise-based undulations
   - Lighting: Set up ambient and directional lights with blue underwater ambiance
   - Water Effects: Implement fog and particle system for bubbles

4. Aquatic Animals (5 Species)
   Clownfish
      - Orange body with white stripes and black fins
      - Orbital swimming pattern at mid-depth
   Blue Tang
      - Bright blue body with yellow tail and black markings
      - Figure-8 swimming pattern
   Green Sea Turtle
      - Realistic shell geometry and flipper animations
      - Slow, graceful swimming near bottom
   Moon Jellyfish
      - Translucent dome with flowing tentacles
      - Vertical floating motion with tentacle sway
   Reef Manta Ray
      - Large wingspan with realistic wing-flapping animation
      - Gliding motion near surface

5. Interactive Controls
   - Drag to rotate: OrbitControls for camera orbiting
   - Scroll to zoom: Mouse wheel zoom functionality
   - Shift+Drag to pan: Pan controls for view movement

6. Environmental Details
   - Seaweed: Create swaying plant models using custom geometry
   - Bubbles: Particle system with upward movement and random drift
   - Water Current: Subtle animation system affecting all elements

7. Animation System
   - Implement useFrame hook for continuous animations
   - Create reusable swimming patterns and behaviors
   - Add realistic physics-based movements

8. Page Integration
   - Update main page to display 3D Aquarium instead of/alongside Space Invaders
   - Ensure responsive design and proper canvas sizing
   - Add loading states and error boundaries

9. Performance Optimization
   - Implement proper disposal of geometries and materials
   - Use efficient rendering techniques
   - Add fps monitoring capabilities

The Results

GPT-5 (Cursor)
Cursor's planning interface was impressive — it broke down my request methodically and asked for approval before starting. But when it came to execution, the result felt more like a proof of concept. Fish animation was minimal, environmental effects were sparse, and while it worked, it didn't capture the aquarium vision from my prompt.

Here's what GPT-5 produced:

GPT-5/Cursor implementation - functional but basic

Claude Code (Sonnet 4)
Claude's approach was less ceremonious but hit more of the actual requirements. While the fish were still pretty basic, it made more genuine attempts to include the species I requested, added more environmental details, and felt like it was actually trying to match the spirit of my prompt.

And here's Claude Code's version:

Claude Code implementation - closer to the prompt requirements (though still imperfect)

Neither result was production-ready, but Claude Code's version made a more honest attempt at fulfilling the specific requirements, even if the execution was still quite simple.

Why It Matters

This wasn't just about building an aquarium. It was about testing whether I'd stick with Claude now that GPT-5 was out.

Here's what I learned:

• Same model ≠ same results — interface, tuning, and workflow matter a lot
• GPT-5 (Cursor) shined in structured planning but felt weaker in direct execution
• Claude Code seemed better tuned for coding tasks that need more follow-through
• For complex technical requirements, Claude made more genuine attempts to implement what was actually requested

Final Take

On launch day, I really wanted GPT-5 to surprise me. And while it's impressive, for coding something creative and detail-heavy like this aquarium, Claude Code still feels like the one I'd trust first.

That could change as tools evolve — but for today, Claude keeps the crown. 🐠

Have you tested GPT-5 against your favorite AI coding assistant? What differences did you notice? The AI development landscape keeps getting more interesting.