Did some stats on context size of agents, broken by role and model.

You can tell that 6.1-sol < sol-6 < astra as stronger model in same role finishes tasks using fewer tokens.