You didn't even try running a model remotely as large as GPT-5. This is like someone running Gemma 270m on a Raspberry Pi and declaring that it's faster than your Macbook.
What I’m saying is that a model not as large out performed in my prompt which is I imagine complex enough to cover a lot of consumers needs. Not to mention, with the right setup, it’d be easy to forward the prompts that need the foundational models through openrouter and only get charged API token prices per use.