A lot depends on whether the z.ai CEO, who just released the first freely-available Opus-class model weights, is blowing smoke when he claims he's less than a year away from achieving Fable-level performance.
If he walks the talk, I really do not understand how either OpenAI or Anthropic is going to justify the twelve-digit valuations they are hoping for. They will just be some people who bought a domain name and rented some GPUs.
I would expect he's 6 months behind. That's exactly how much lag the open models have versus the frontier models, and have consistently the last 2-3 years.
The question is, how far ahead will the frontier models be in 6 months? if it's still 6 months, open weights might have a fable equivalent model, and the frontier models will be on upwards towards ... essay, or novel, or bibliography, or whatever the next name is.
After a point, does it matter? Even assuming that models can keep getting smarter indefinitely, it's not like you need to use the smartest model to get the job done.
Moore's law is dead now, so at some threshold purchasing the GPUs to run the biggest and newest model hurts you more than whatever rent you could've extracted from it.
When we get there, why would you want to run a closed model that you can't control, with restrictions you can't remove, that a company can take from you or silently nerf without telling you?