Context tokens are not directly based on the model unless talking about very large token counts. Context tokens come from the amount of memory it is given to work with, which costs more to host. Some models (not any of the Llama models) are better at managing this memory, but it still uses a lot of resources to host a large token context.