I assume this has to do with the $20 tier now running out of provisioned tokens so quickly as to be not particularly useful, giving users a bad experience.

The million token context + reduced caching period + new models using more tokens made this a probably unpopular but perhaps unavoidable development.

There's a hard problem here balancing costs and experience. I'm afraid despite the bad experience for people that this is necessary and $20/month was just too big a loss to sustain.

> $20/month was just too big a loss to sustain.

Is there any marginal cost associated with a new subscriber?

I have always heard inference is cheap and the cost was in training, so I assumed any subscriber was making them money, just not enough to cover their insane fixed costs.

But I am just guessing.