Model Selection & Token Efficiency
In this workshop, we cover what tokens are, what burns them, and how to pick a model without slowing down.
Key topics covered:
Tokens:
You are billed for tokens in and out of the model, not for agent actions like search. The harness wraps the model and is tuned per model.
Four token types:
Input, output, cache write, and cache read. Output costs more than input. Cache reads cost less, and a rule or model change can rewrite the cache.
Context:
The harness feeds context back each turn. Near 90% it compacts the middle of the chat, not the prefix or your latest ask.
Model choice:
Explore in A