Seems like a pretty bad business move if it's really what they're doing. They should want devs using the product on a cheaper subscription to see the value with profitable limits on usage.

I think the only reason to do this would be that they just can't scale up to service the volume they have and need to cut down significantly on the total number of users. Seems also like a rough business proposition. Most of the pro plan users would probably migrate to a competitor at a similar price point (I know I will).

The only other possibility would be if they are losing too much money on the compute power and just can't offer it at that price anymore. But then upgrading the plan gives you more compute per dollar, so maybe they're just banking on people not actually using all of what they pay for?

I had previously thought that the inference cost of using a trained model was relatively low and that most costs went into training new models, but maybe that is less true with the more powerful newer models.

If it costs a ton more to serve Opus vs serving something like Kimi or Qwen, then I think most people just won't use the more expensive version for most things.