As title suggest, some models from OpenRouter are great for storytelling, but are also expensive.
For example, using a big and complex world, trying Full Grok model seems to be the best performing for me (actually it and Claude Sonnet 3.7 are the only two models I accept, any other models I've tried are just too dumb for my complex setup). But they both are costing me around 0.08$ per request, which can pile up quickly!
But that's because they are being called 3-4 times per request, while the only important one would be the Game Text Prompt. If I could call other cheap or free models for the rest of calls, costs would go down to a much more acceptable 0.01-0.02$ per interaction.
Also the best models for writing are not always the best at following instructions. They tend to be verbose and not so good at maths.
For both reasons, beeing able to call for different models for the 4 different AI interactions would be ideal. Currently there are 4 interactions per game's request if I understand ir correctly:
- Game Text Prompt: This requires a complex model, capable of emotional nuances and extremelly good at detail retention.
- Choices Prompt: This could be done by a general smart model. Paid expensive models don't seem to offer better options than free available ones such as Deepseek or even some Qwen. And the player is likely to override the answers anyway...
- Stat Updates Prompt: This requires an Instruct model better. Good models for writing tend to use extreme values all the time. A small and Instruct model looks like the best choice.
- Location Change Prompt: Is this ever used? I usually change location "manually" by the input, but if it is used, again, there's no need for a big model for this. The same Instruct model could do the task.
Please consider adding this option as it would make your game much more viable, right now I have just dared to build the world prompt and test intensively that it works, but cannot really use it for fun given the costs.
Thank you.