This would be interesting to try with local models, where the token costs and token limits are quite different.