Playing the meta definitely requires opponent modeling, but I don't think it's exclusively opponent modeling. In the example of Splendor, I think meta is best explained as the interaction between emergent strategies. An AI may not need to think of it as a meta as is the case in chess, but an AI may still be playing strategies that match elements of the metagame.

I think European style board games are great examples. Euro board games are often designed in a way where there are 3-6 different "currencies" in the game that are vaguely interchangeable in terms of progress towards the victory conditions, but these currencies have situational tradeoffs. It's usually a safe bet to play a T-shaped strategy where you focus on a few currencies but dabble in the rest enough to influence the game conditions to favor you. Scoring mechanisms often explicitly reward being T-shaped and gameplay mechanics often explicitly favor T-shaped strategies, so regardless of whether the AI is explicitly trying to play a meta strategy, it's probably going to look like a meta strategy because the meta is T-shaped.

By my definition it's meta to do opponent modeling since you are using information outside of the game mechanics.

And you're right, for games with high efficiency and public knowledge you can often just grind through optimal play.

However, for games with hidden information, you can play theoretical optimally, but you can also gain an additional edge if you can properly model your opponent. I'm thinking like in Poker. There's mathematically correct play, but because you play against folks who don't play mathematically correct you could get an extra edge by properly modeling the faults of your opponent you can exploit!