Hit this myself recently, along with a bunch of overloaded errors. I think it's growing pains for where we are with AI right now.

As the tooling matures I think we'll see better support for mixing models — local and cloud, picking the right one for the task. Run the cheap stuff locally, use the expensive cloud models only when you actually need them. That would go a long way toward managing costs.

There's also the dependency risk people aren't talking about enough. These providers can change pricing whenever they want. A tool you've built your entire workflow around can become inaccessible overnight just because the economics shifted. It's the vendor lock-in problem all over again but with less predictability.