Cost-Aware Routing

Route based on difficulty.

The Cascade

1. Try Llama 3 8B (Fast, Cheap).
2. If answer confidence < 0.8:
3. Try GPT-4o (Slow, Expensive).

The graph logic handles this fallback automatically. LlamaNode -> CheckConfidence -> GPT4Node. For 80% of tasks, you pay the Llama price. For the hard 20%, you pay for intelligence.