It does seem to be that way, though, I'm very curious to what capacity and how they are using Anthropic.

I suspect they just have loads of subscription users, probably many enterprise seats, but I sort of doubt there are agentic pipelines built on their metered pricing in the same scale as they may have subscription seats.

It also depends on how much compute you want to devote to training. If Anthropic is training heavy and compute constrained, users will feel it.

I think the whole pentagon thing really cast Anthropic into the spotlight, and their aesthetic is pure marketing excellence. It is the "trendy" AI to use. However on the backend, they did not have the compute available to serve everyone, especially since they seem to lean towards massive parameter models.