It's also possible they have found a way to run the model extremely cheaply. To be fair, there has been many improvements to transformer inference since they in

It's also possible they have found a way to run the model extremely cheaply. To be fair, there has been many improvements to transformer inference since they initally set their prices (most notably flash attention), so if they were barely making a profit back then they could still be making a profit now. That's a big if, however, and no one really will give you figures on exactly what this costs at scale. Especially since we don't know for a fact how big GPT-3.5-turbo actually is.