Maybe play around with CPU inference on smaller models too. Frontier models are probably overkill for simple language responses at this point; I suspect you could get pretty good results with Qwen doing this, and Kokoro TTS is really good now.

Instead of that, I would not put in my own key, you should get a brand partnership. Something like the Liquid Death Murder Your Thirst water brand. https://liquiddeath.com This would be perfect for them and paying for all the AI use would be nothing compared to their normal advertising. Win win.

Are you proposing a small add on the site? or does the AI agent do something like this:

"Hey give me a second to answer while I take a sip of this crispy liquid death, these killer cans of soda-style sparkling will attack your tastebuds with soda-level flavor and only 1/10th of the sugar."

"Please slurp as loud as possible into the mic"

"Drink verification can to continue"

Dude, just wanted to be sure you saw his comment about exposing your API key. You do not want to do that!

I think I replied to this, ephemeral secrets are minted per browser