So far OpenAI's modus operandi was "does it get better if we make it bigger". But now we've reached "good enough" (arguably already with GPT3.5, as evidenced by free ChatGPT running on GPT3.5-turbo).

There has been promising work about making models smaller without losing performance (e.g. by training them longer), or quantizing the weights (llama.cpp etc) to make the model cheaper to run. I believe we will see a lot more focus on this over the next couple years, and given how there's been comparatively little work done in this direction I wouldn't be surprised by 10x or 100x efficiency gains.