Simplify AI picking dropdown

[novoselt]

I specifically was thinking about but it applied to other cases as well.

Critique of the current state - it shows some information, but on the one hand it feels like too much, if all I want to do is change a model and I already know what to pick, while on the other hand it is too little to make a decision if I am not familiar with the models.

In particular "Free" is unambiguous, but what does "Paid" mean??? There is no price shown at all and we can't just show "$0.12" for one and "$0.04" for the other. So instead, I think there should be a dropdown JUST WITH NAMES and a button next to the dropdown "Help me choose..." which opens a page with

• recommendations on what system to choose
• exact prices
• "typical" price - I think this is hugely important, tokens are a very abstract thing, how much will I be charged if I ask to generate a cell? $5 or $0.02? Perhaps we can show something like "average/median cost per request on CoCalc during the last month" - that would be a good starting number.

The current price table in purchasing limits is very difficult to find. I suspect that "organic discoverability" for it is about 0. It would be also much easier to interpret if it was in cents per 1k tokens, not dollars, which requite one to carefully keep track of zeros.

[williamstein]

This is not happening in a vacuum. It's very relevant to consider how all of the models advertise their own pricing. The current format (e.g., dollars per 1000) is copied exactly from OpenAI from a few months ago. OpenAI changed their pricing lists recently to be dollars per million tokens. https://openai.com/pricing#language-models

Anthropic gives pricing in dollars per million tokens (instead of 1000): https://www.anthropic.com/api

Most other models now give price in dollars per million tokens. They are all very careful to distinguish between input and output, of course.

tokens are a very abstract thing

It's not much more abstract than "word". It's roughly 1.3 tokens per word in English.

[novoselt]

That is definitely true and I've seen those tables. I suspect, however, that most users are unaware of them, so something that makes sense for app developers is not necessarily the best presentation to users.

For tokens conversion to words is still not very useful, I think - the question is "How much I will be charged if I start using it?" and some idea of "5 cents per request" is way more clear that "here is the formular that will give you an exact number for each particular use case, all you have to do is to count the number of words in your document and take into account our prompt". Given average/median for other users does not come with any guarantees, and we certainly should make it clear, but in most cases it will be a very accurate reflection of use experience, I suspect.