Given it is being trained on your project, the latter isn't that surprising. For the former, you could use LLMs yourself for the probabilistic matching as alternative method? Probably you don't because the trade-offs (like performance) are not worth it...
Yes - it's certainly the case at the moment that you can run a few thousand through the LLM at a reasonable price, but not, say, ten million. But the rate of progress suggests to me that this argument won't hold up forever. Eventually I think an off the shelf LLM will outperform most and probably all more traditional ML models at this task. Largely because LLMs can identify tricky ones and pick them out for more intensive effort (e.g. looking online, further searches again the canonical list of addresses)
I didn't check which model exactly you are using, but I think a complex LLM, that is capable of deciding to look online, will probably always be more expensive than most classical models. Maybe if someone invents a way that reliably strips every other ability than answering the one question one has and checking online sources, the LLM can reach an equal level.
