If GPT-4 arrives, and the model size grows by as much as many hope (1 Trillion+ parameters), then it would likely cost more to run than GPT-3 would. So even if GPT-3 prices would come down by half, then I would expect that GPT-4 could be double what we’re paying now for GPT-3… unless they somehow find a magic way or running it less expensively. Everyone talks about the much faster exponential growth of AI relative to Moore’s law, but if you look at the data, the main contributing factor is the willingness to invest more money in AI solutions… the exponential growth of dollars is a bigger factor than hardware or algorithms. So a bigger GPT-4 would likely be expensive but still worth every penny 