5 ms·

for startups I guess this means nail your use case with gpt-4, and when scaling cost becomes an issue consider fine tuning.