I batched 150mb of textual data to extract data from as prompts to OpenAI to analyse the data and give me facts about it back to me in json format. To give you an idea about how much 150Mb in text is, the first Harry Potter book is 429Kb.

Ok people, how much do you guys think it costed? I used gpt4o-mini model. The result will shock you.

5$?

idk

@jestdotty wow actually. Will tell you later how much wrong you were after others gave it a try.

@retoor 69 cents

@Demolishun or 4.20 canadian rubbles

80 cents? Maybe $1.20

π €

Considering your whole article on molodetz about how cheap AI, I assume some comically small amount like, idk, 29 cents or something

Okay, went back to your article, and then visited gpt's pricing for 4o-mini, now my answer is $4.5 (Not changing my original answer because that's half cheating since that's against the spirit of guessing), but that's just for input, you gave no indication for the scale of the output so no way to guess or measure it.

around $7

I'm surprised no-one has said 3 fidy

I'm surprised but it was 20,-. Better guesses than expected! But why am I sharing this? Many people go for self hosted solutions for Llm's and so did I at the beginning. But with such pricing, why would you ever do local?

What is the data? All my keyboard input of the last few months. Yes, I wrote 300 harry potter books in amount of keypresses.

Another about self hosting vs hosted: local llm's are woke too.

But also a realization: if you pay for got the 24,- a month.. Realize that you pay for 300 harry potter books on input and in reality won't do 25% of one book. But the api access and use openwebui, not Chatgpt itself

I'm very happy with your realistic estimations!

@retoor Because I'm a niny and I like being able to see it work ;P

Depends on your data format.

You appear to use english as a primary language, but we don't have enough information to know the languages of the source text, and its unlikely the source text was generated by you.

So we have to make some big assumptions to answer your question.

Assuming ASCII, or UTF-8 text, primarily english, we assume one byte per character.

We also don't know the format the text data is in, so it actually may be using chars not in the ascii range of utf-8 (first 128 characters or first byte, correspond to the first 127 in ascii).

We're looking at 157,440,000 bytes IF our assumptions are correct.

Assuming o4 mini generates as many characters out as it takes in input, and because it is a reasoning model (which tend to saturate their reserved output tokens), any given run should use the entirety of the tokens available for that run (this is a bit of guesswork, but for the purposes of napkin mathing it, it will be alright).

Now, with 157.44 million characters, and an average of 4 characters per token (a simplification due to BPE encoding), we're looking at 39.36 million tokens to process.

39.36 million tokens input, times $0.15 per million tokens, equals $5.904 for the cost of the entire input.

And assuming it generates at least as many tokens output (including reasoning tyokens) per input (discounting the separate cost of reasoning tokens, which I'm unsure of), factoring that it charges $0.075 per million tokens out, you're looking at $2.952 for the output of the entire set, for a grand total of $8.82 to process your entire data set.

Was I close?