Open Source vs frontier models on a single-file HTML canvas driving animation – results

Hey yall, I was inspired by this post : https://www.reddit.com/r/LocalLLaMA/comments/1tf3p6c/local_qwen_36_vs_frontier_models_on_a_coding/

And I know this isn't exactly local, but I wanted to share what I tested out and what results each model delivered so I decided to share this.

I ran the same single-file Canvas prompt across multiple models using my harness
( https://github.com/AidenGeunGeun/OpenCodeOrchestra ). The models were able to use whatever tools they had access to - some used auditor models and some did not. We have some clear winners and some ambiguous.

https://preview.redd.it/2ehkh47vfo1h1.png?width=2972&format=png&auto=webp&s=d1e643f7b8bd0c3bab241838731848109359f1e1

The results are here:

https://aidengeungeun.github.io/oco-canvas-car-scene-compare/

Setup:

• Same prompt for every run
• One isolated Orchestrator per model
• Highest available thinking/effort setting for each model
• Output target: one standalone HTML file, no libraries, no external assets
• Task: realistic side-view car driving scene with parallax scenery, spinning wheels, subtle body motion, cinematic lighting, and seamless looping

Models included:

• GPT-5.5 xhigh
• GPT-5.4 xhigh
• Claude Opus 4.7 (max effort)
• Claude Opus 4.6 (max effort)
• Claude Sonnet 4.6 (high effort, max doesnt exist on sonnet)
• Kimi K2.6
• DeepSeek V4 Pro
• DeepSeek V4 Flash
• GLM-5.1
• MiniMax M2.7
• Qwen 3.6 Plus
• Grok 4.3

I used whatever highest thinking possible for each model. tok/s and time for generation were not measured.

The results are here:

Gallery: https://aidengeungeun.github.io/oco-canvas-car-scene-compare/

Source: https://github.com/AidenGeunGeun/oco-canvas-car-scene-compare

We know that models are capable of doing these kind of work, but I was wondering how a wide variety of Open weights models compare to frontier models, especially the ones that are used often.

I tried to use MiMo-V2.5-pro too, but since that model had billing issues with the OpenCode Go subscription, I couldn't use it. Take a look!