Network effects: hidden and in plain sight
Social networks tend to be free to use, because every additional user strengthens the system. A paid tier with exclusive features often don’t work because holding back features makes the free product less useful, and every user lost is very bad for the network.
Products that are built with machine learning are similar, except the network effects are hidden. The data generated by each user can be used to make the system better, and, in general, whoever has the best data, wins.
A paid-only subscription to one of these products will typically result in a worse product than a freemium model, and it’s likely that a totally free model will be even better. The exception is when an important group of users feels compelled to pay to get the best product. For example, business users purchasing software. In this case, freemium is the way to go.