FIELD BENCHMARK · 5 FRONTIER MODEL RUNS RUN ID: 1f7a9135-4696-4176-9bd0-b91c9ce970f9

One whimsical prompt.
Five miniature 3D worlds.

We gave Google Antigravity (Gemini 4 Carbon), ChatGPT Codex (GPT-6.2-Astra & GPT-6.1 Sol), Claude Fable 5.2, and Kimi K3.1 Pro the exact same open-ended prompt: build a cute, funny-named 3D traffic simulator busting everyday road myths. Play all five live below—and inspect where clean engineering, aesthetic taste, dual-renderer craft, and over-testing diverged.

Launch Live 3D Arena ↓
MOST COMPLETE & CLEANEST CODE 16:00 · 13 Labs

Antigravity (Gemini 4 Carbon) built 2×–3× more experiments than any rival with the cleanest Firebase-ready ES module architecture.

TWO CODEX GENERATIONS GPT-6.1 Sol & 6.2-Astra

6.1 Sol (`16:20`) nailed the cutest warm editorial UI (`6` labs); 6.2-Astra (`~13:00`) sculpted a rooftop-donut-shop town with baked WebP sprite fallbacks.

THE MONOLITH QUIRK 1 File · 66.3 KB

Claude Fable 5.2 (`35:00`) deployed straight to cloud with its signature right-hand side panel—but crammed CSS, UI, and 5 Three.js scenes into a single `.html` file.

THE TESTING PARADOX 54:00 · 2 Broken Labs

Kimi K3.1 Pro spent the longest verifying UI & optimizing mesh triangles—yet cars ghost straight over each other in its roundabout and its speeding timer goes unread.

INTERACTIVE SANDBOX THEATER

Test All Five 3D Simulators Live

GALLERY & ARCHITECTURAL CRITIQUE

Inside All Five Builds

Click any card’s Load on Stage or Fullscreen button to test its physics, camera controls, and UI ergonomics directly.

CROSS-MODEL PHENOMENON

Why Did Five Independent Runs Build Almost the Same Toy Town?

Nothing in the prompt specified a rendering engine, car proportions, loop topology, traffic math, or color scheme. Yet across five completely separate runs, the outputs look like siblings from the same miniature studio:

  • 100% Procedural Three.js Chonk-Cars: Zero external `.glb` models—every LLM built rounded pastel vehicles with brake-light emissives, cottages/donut shops, and lollipop/pine trees in code.
  • Alliterative Candy/Sound Naming: BeepBeep BeanTown!, Honk Hollow, Jammy Little Roads, Jellybean Jam, and Bumper Buddies.
  • IDM Ring-Road + 4 Universal Labs: Every run reached for the Intelligent Driver Model on a closed stadium/circle loop, plus dual-lane passing, roundabout vs. crossroads, and a braking/kinetic-energy speeding breakdown.
WALL-CLOCK BUILD TIME (MINUTES · LOWER IS FASTER)
ChatGPT Codex · GPT-6.2-Astra ~13:00 · 4 labs
Antigravity · Gemini 4 Carbon 16:00 · 13 labs
ChatGPT Codex · GPT-6.1 Sol 16:20 · 6 labs
Claude · Fable 5.2 35:00 · 5 labs
Kimi · K3.1 Pro 54:00 · 5 labs

More wall-clock time did not buy fewer bugs: Kimi spent 54 minutes running UI and triangle-count passes while leaving two core traffic simulations broken.

SOURCE-CODE FORENSICS

Under the Hood: Clean Modules, Baked Fallbacks, Monoliths & Bugs

SIDE-BY-SIDE BENCHMARK DATA

Engineering Scorecard & Scenario Matrix

Head-to-Head LLM Coding Benchmark Scorecard (Experiment 1f7a9135 — 5 Runs)
Model & Harness App Title Build Time Labs Built Code Structure Cloud & Assets Design & Physics Verdict
Google Antigravity Gemini 4 Carbon (Medium) BeepBeep BeanTown! 16:00 13 Labs 19 ES modules · 10,954 LOC (cleanest separation) Firebase Hosting config ready; mobile needed +1 prompt Most ambitious traffic science coverage; rock-solid roundabout & Braess physics
ChatGPT Codex GPT-6.2-Astra Honk Hollow ~13:00 4 Labs 7 ES/CSS files · 1,081 LOC + baked WebP sprite fallback OpenAI cloud + 3.46 MB offline single-file export; zero CDNs Richest 3D town props (donut shop!), twin synchronized junction sims, fixed views
ChatGPT Codex GPT-6.1 Sol (Medium) Jammy Little Roads 16:20 6 Labs 5 files · 404 LOC + `node --test` unit suite Deployed to OpenAI cloud out of the box; responsive Cutest & cleanest editorial UI; free-orbit cam, honest citations & clean IDM
Claude Fable 5.2 (Medium) Jellybean Jam 35:00 5 Labs 1 monolithic `.html` · 609 lines (hard to maintain) Deployed to Claude cloud out of the box; bottom-sheet mobile Typical Claude side-panel look; fun 16× warp & Tailgating bonks
Kimi K3.1 Pro (Medium) Bumper Buddies 54:00 5 Labs 4 ES modules · 2,412 LOC + Three.js chunk split Local portable `dist/`; heavily tested responsive layout Most QA & triangle optimization passes, yet roundabout & speeding physics failed
Traffic Phenomenon Coverage Across All 5 Runs (14 Total Scenarios)
Traffic Scenario / Paradox Tier Antigravity (13) GPT-6.2-Astra (4) GPT-6.1 Sol (6) Claude (5) Kimi (5)
EXPERIMENT DOSSIER // 1F7A9135

Prompt & Benchmark Methodology

EXACT INITIAL PROMPT

“Create a 3d car traffic simulator, showing different traffic cases, like how traffic jams are formed, does the other lane really move faster or it's an illusion, do roundabout cause less accidents then intersections, does speeding actually help you get to your destination faster, by how much and how much it increases your risk on the road, and so on... Come up with a funny and cute name for it, and make the graphics cute and engaging”

Models Tested

  • Google Antigravity — Gemini 4 Carbon (Medium) · 16:00
  • ChatGPT Codex — GPT-6.2-Astra · ~13:00
  • ChatGPT Codex — GPT-6.1 Sol (Medium) · 16:20
  • Claude — Fable 5.2 (Medium) · 35:00
  • Kimi — K3.1 Pro (Medium) · 54:00

Key Observations

  • Antigravity built 13 labs with the cleanest modular Firebase-ready code (mobile took +1 prompt).
  • GPT-6.2-Astra built the richest 3D props (donut shop!), zero-CDN self-hosted fonts/Three.js, synchronized twin junction sims, and a pre-baked WebP sprite fallback.
  • GPT-6.1 Sol nailed the cutest, cleanest editorial UI; Claude used its classic side-panel look in a single HTML file.
  • Kimi invested 54m in UI/triangle QA, yet cars overlap in its roundabout and its speeding experiment fails.
LIVE 3D SANDBOX

Simulator

Open Standalone Tab ↗