Maybe the model doesn't do multiple queries but just one long query guided by thought tokens.