Level 1.
Right now, there's not really any easy space in most classes to implement a subclass that messes with the base assumptions of the class re ability scores.
WotC wrote the Hexblade to have a way of implementing a melee-focused warlock without necessarily requiring that warlock to have a very high Str or Dex. They created a subclass that used Cha as the melee stat, but they were able to do that functionally because warlocks take their subclass at level 1. And this sorta worked.
For instance, if i write an ironskin monk subclass, a subclass focused on absorbing or resisting damage rather than evading it, that keys some of its monk abilities off Str and Con rather than Wis or Dex - a PC planning to take that class will presumably put higher scores into Str and Con. But that means they're going to be crippled until level 3, relying on sub-par Dex and Wis scores until they gain the first subclass ability.
Same if I want to implement a whirling dervish type barbarian subclass, that emphasises speed and the use of light Dex-based weapons rather than Str. If I want to have a subclass ability that allows the rage damage bonus to apply to Dex weapons (pretty fundamental for a class like this), that ability won't apply until 3rd level. The PC will be badly hamstrung in the early game (which is where PCs are already most vulnerable).
You can see a bit of the same going on with the armorer artificer. At level 3 they get proficency with their special armour, which will normally be heavy and not allow dex modifiers to AC. So in the long run, putting a high ability score into dex will largely be redundant. But if the artificer does this, then for the first couple of levels they're stuck being extremely vulnerable with an extremely low AC.
Personally I'd like a bit of a rewrite of most classes to be more flexible about which ability scores they key off. An int-based, low-cha warlock is almost a classic archetype of the sort of person who would make a warlock pact, plus I think there's plenty of room for wis-based sorcerers, big brutish str-based rogues etc etc. But that seems distinctly unlikely to happen, so in the meantime, more space for convention-breaking subclasses that can do the job instead, please.