Cost for higher token usage

[cralicriux]

Let me start off with, this project is amazing! I'm still very new to Claude code and PAI. I've been using opencode for a few weeks for AI in general, dipping my toes into an AI assistant, but with Opencode I haven't had to see how much tokens I'm using really.

But let's get to the point. It seems that, and makes sense too, that PAI is using much more tokens. Has any thought been put into having different models/modes to have PAI (automatically determined based on the tasks) run in to better keep token usage down? I've already burned through my Claude pro tokens each day within about an hour :P

Right now I'm thinking to manually just switch the model as I'm using PAI, but thought it would be cool to have PAI set up so that certain low level functions use lower level models to save costs/time.

[danielmiessler]

Great question! The token costs depend on:

1. Model used - Opus is most expensive, Haiku is cheapest, Sonnet in between
2. Context length - PAI skills add context but are designed to be efficient
3. Tool calls - Each tool invocation adds tokens

PAI v2.1 focuses on token efficiency:

• Skills load only when needed (deferred loading)
• Browser skill uses file-based MCP pattern (99% token savings)
• Minimal context injection by default

For high-volume usage, consider using Haiku for delegation tasks and Sonnet for main work.

See: https://github.com/danielmiessler/PAI/releases/tag/v2.1.0

[danielmiessler]

Short version: use Sonnet for most tasks and reserve Opus for deep Algorithm runs (Extended effort level and above). PAI's effort level system (Fast/Standard/Extended) is designed to match model capability to task complexity, which keeps costs manageable.