> I know there are open source alternatives, but they’re not nearly as capable

They certainly are capable (DeepSeek being the obvious example), the problem is that they're still too expensive to run and there's no currently differentiator to compete with the big players who are likely selling inference at a loss.