> price per token, which has fallen.
I think that statement is incomplete. It's cheaper because the AI providers are subsidizing the queries by burning cash in order to gain market share. So it's not a price, it's a subsidized and temporary price. Which will likely go up once burnable money runs out / the providers switch from "market acquisition" to "let's try to make this thing profitable at all".
Also, I'm sure I read something here in the last few weeks documenting that while every specific model shows falling price-per-token, "frontier models" which are the ones everybody wants to use - maintain steady prices (and are often significantly more expensive to use because they consume way more tokens than older models responding to prompt-fashion-de-jour).
I'm not sure this is true anymore. Anthropic e.g. claims that each model is profitable, seen over its lifetime. There are also a ton of providers of open-source models that are popping up that are hosting models for very cheap prices per token - hard to say how subsidised they are, of course, but I would be surprised if they are all burning cash.
Open-source models are also constantly getting better, with new architectures and training processes that make 30b models perform like previous SOTA models, lowering the prices for many use-cases.
So I don't think the current level AI will get more expensive, I would expect it to stay at the same price or fall. But, I think the AI companies will continue pushing more higher-priced premium models that deliver better performance for higher prices and with more tokens consumed.
This is already happening. Some time ago all models providers charged $20 a month. Now you still get $20 version, which is crippled, with limits. If you want "a real thing", you need to pay $200. And this might still be too little, as the algo that runs the show is quadratic with the size, which, as every algos books says, is highly inefficient.
[dead]