You pick it at level-up, not at creation, and with only 12 levels to spend, every one you put into a second class is a level your first class doesn't get — which is exactly why short "dips" work better than even splits.
You multiclass at level-up, not at character creation: instead of advancing your current class, you pick a second one to put that level into, and you can keep mixing from there. There's no ability score requirement to do it in Baldur's Gate 3 — Larian dropped that tabletop rule entirely — but the level cap is only 12, so every level you hand to a second class is a level your main class doesn't get, and that's the real cost.
On the level-up screen there's a separate control for choosing a class — a small crossed-icon button next to the usual "continue this class" option. Pick a different class there and that level's worth of hit points, proficiencies, and features come from the new class instead of the one you started with. The first time you dip into a class you also pick up its weapon and armor proficiencies, plus one skill if the class is Rogue, Ranger, or Bard. Each class then tracks its own level separately — a Fighter 3/Cleric 1 has exactly the powers of a level-3 Fighter and a level-1 Cleric, not some blended level-4 hybrid.
Tabletop Dungeons & Dragons goes to level 20, which is enough room to run three or four classes and still finish strong in each. Baldur's Gate 3 stops at level 12, so the same three-way split that's viable on paper leaves every class shallow here. Two levels spent on a class you're using for one feature is two levels your main class isn't spending on its own progression — with only twelve total, that trade is felt immediately, not smoothed out over a long campaign.
The features you give up by not staying single-class tend to be the big, level-gated ones:
The reason multiclassing is still worth doing is that one or two levels in the right class buys a specific, load-bearing feature without costing much of your main class's progression. A few combinations come up often across build guides for exactly that reason: a level or two of Fighter grants Action Surge and heavier armor without derailing a caster; two levels of Warlock into a Bard adds Eldritch Blast (and Agonizing Blast to scale it with Charisma), covering a Bard's usual lack of a strong damage cantrip; a single level of Cleric ahead of a Wizard can add heavy armor proficiency depending on the domain; and two levels of Paladin grafted onto another Charisma-based class unlocks Divine Smite for burst damage on weapon hits. The pattern across all of them is the same: take just enough levels to grab the one feature you want, then go back to leveling the class that does the heavy lifting.
In pen-and-paper 5th edition, multiclassing into a new class requires meeting a minimum score in its key ability — 13 Charisma to pick up a level of Warlock, for instance. Baldur's Gate 3 removes that requirement outright: you can multiclass into any class the moment you have a level to spend, regardless of your stats. It's a real convenience, but it also means the game won't stop you from taking a dip that doesn't actually work with your build — that judgment call is still on you.
Multiclassing isn't permanent. Once you've met Withers and brought him back to your camp, he offers a full respec for a flat 100 gold, any number of times. A respec rebuilds class, subclass, levels, ability scores, skills, spells, and feats from scratch, so a multiclass spread that isn't working can be undone completely rather than lived with. What it won't touch is your race, background, name, appearance, or anything tied to quests already resolved — those choices, unlike class, are the ones worth getting right the first time. If you're still deciding what to build before you even reach the character sheet, that's covered in more general terms in our guide on what to know before you start — and if a companion like Karlach is already in your party, her ending has nothing to do with your class choices, but it's the other question people tend to have open at the same time.