I Tested ９ AI Models in a Real Scenario (Surprising Results)

The Test

I gave 9 AI models the same coding challenge: write code to generate an animation. Same prompt, same constraints, same evaluation criteria. No cherry-picking, no re-prompts — just raw output.

Models Tested

#	Model	Provider
1	GPT-5.5	OpenAI
2	Claude Haiku	Anthropic
3	Claude Sonnet	Anthropic
4	Claude Opus	Anthropic
5	Gemini 3.1 Pro	Google
6	Qwen 3.6 27B	Alibaba
7	Kimi K2.5	Moonshot
8	GPT-5.4 / Mini	OpenAI

Full Results

All prompts, model outputs, and my notes are available on the companion benchmark page:

📊 kuro-llm-benchmark.pages.dev

What I Learned

• Different models have wildly different strengths — the "best" model depends on the task
• More parameters don't always mean better output
• Real-world testing reveals things that benchmark scores hide

Watch on YouTube

Full video with panda commentary: youtube.com/watch?v=CeMXQGfNuXo