Build/vibe-code your own benchmark and find out, the work pays off and lets you try many models.

Also try fine-tuning small models and see how they perform.

Oh, I have, and even Opus falls flat. I feel like regex would be better. Oh... maybe i can have Opus create such a regex...

If you haven't yet, you might want to try gemma-4-E4B-it-qat or gemma-4-12B-it-qat with Structured Outputs. My main use case is tagging photos and generating headlines.