Is Claude 3 Sonnet overkill for simple product descriptions?

I’ve been testing the new API limits for our e-commerce catalog refresh, and I’m hitting a wall with cost optimization. We have about 5,000 SKUs that need basic 150-word descriptions for a mid-tier home goods store. The content is pretty formulaic: material, dimensions, use case, and a slight lifestyle hook. Previously, I used GPT-4o-mini and it worked fine, but the tone felt a bit generic. I switched to Claude 3 Sonnet to see if it could inject more personality without becoming verbose. The quality jump is undeniable; the copy sounds less like a robot and more like a human who actually cares about the product. However, the token usage is significantly higher than the mini version, and our margin per item is thin. Has anyone else run into this trade-off? I’m considering a hybrid approach where I use the cheaper model for the factual specs and only prompt Claude for the final lifestyle sentence. But I’m worried about inconsistency in voice if I’m stitching two different models together. Also, does anyone have a specific prompt strategy that forces Claude to stay strictly within word limits? I keep getting responses that are 220 words long, which breaks our UI layout. We’re using structured output JSON, but the model seems to prioritize flow over the hard character constraint. Is there a way to penalize length in the system prompt effectively, or am I better off just truncating in post-processing? Looking for real-world experience here, not just theoretical best practices.