Give engineers unlimited tokens at the start. Don't optimize early. A less capable model often uses MORE tokens to complete the same task — so the 'cheaper' model actually costs more. Use the best model and let people explore freely.