Maybe they do that opaque degradation trick that whenever it's asked something questionable, it'll route to a worse model instead.