I am a little bit confused about the pricing (price per 1M token) regarding the gpt-4o model.
OpenAI lists the price for 1M input Tokens with $5 for “gpt-4o” model and with $2.5 for “gpt-4o-2024-08-06” which seems to be the lastest model.
The $5 price is also listed for the older “gpt-4o-2024-05-13” model. (See screenshot)
I thought that “gpt-4o” always points to the latest model (as the API documentation mentions).
So which pricing is used when “gpt-4o” is used and which model?? Does it really become cheaper if I would use “gpt-4o-2024-08-06” instead of “gpt-4o”?
Yes, it would be cheaper currently to use gpt-4o-2024-08-06 compared to gpt-4o. Ideally, applications in production should directly use the dated snapshot.
gpt-4o-2024-08-06 has several changes that could be performance-breaking, so you should pin to a model version anyway. It is built on “instruction hiearchy”, a trust model that can affect the AI belief in a prompt an application is supposed to adhere to. That may be why the alias hasn’t been pushed up as rapidly as others.