Presumably the hosting services are resolving all of this in their OpenAI/Anthropic compatibility layer, which is what most tools are using. So this is really j

Presumably the hosting services are resolving all of this in their OpenAI/Anthropic compatibility layer, which is what most tools are using. So this is really just a problem for local engines that have to do the same thing but are expected to work right away for every new model drop.