New figures show that if the model’s energy-intensive “chain of thought” reasoning gets added to everything, the promise of efficiency gets murky.