frozen=True overrides __setattr__ to raise FrozenInstanceError on every assignment after construction. As a side effect Python also synthesizes __hash__ (mutable dataclasses are unhashable by default), which makes the instance usable as a dict key, a set member, or an lru_cache argument. Use frozen for value objects like coordinates, ranges, money amounts: anything where two instances with the same fields should be the same logical thing. The cost is dataclasses.replace(c, x=99) instead of c.x = 99 for updates.

Without slots, every instance carries a per-instance __dict__ so you can attach arbitrary attributes at runtime. slots=True declares the exact set of allowed attribute names, drops the __dict__, and cuts memory by roughly 40% per instance. The trade-off is that typos like n.scope = 1 (when you meant n.score) raise AttributeError instead of silently creating a new attribute, which is usually a feature, not a bug. Reach for slots when you allocate millions of instances of the same class (graph nodes, ECS components, parsed records) or when you want the class to refuse mystery attributes.

frozen=True, slots=True is the canonical 'value object' shape: immutable, hashable, memory-flat. frozen guarantees the object cannot drift after construction (no Money instance ever changes its amount), and slots keeps the per-instance footprint tiny when you have lots of them. Updates use dataclasses.replace(obj, field=new) which produces a new instance with the field changed, leaving the old one untouched. The combo is the right starting point for domain primitives in any non-trivial codebase: prices, intervals, identifiers, version numbers, geographic coordinates.