How much do OpenAI and LLM AIs Cost These Days?
I did a little cost analysis of different well-known AIs! Ran a bunch of conversational prompts and tracked the cost.
Price data for OpenAI from https://openai.com/api/pricing/ as well as from actual billing, I wanted to check their accounting!
t = token, M = million, q = query
All based on Uma-Y 1.8 using standard biographical sketch (details below). Results apply to other GPT/LLM AIs in devices like ConvoBot and Sub Controller.
OpenAI gpt-4o $2.50/1Mt input, $10.00/Mt output
15 q, roughly 2k t/q, around 35635 t
expect bill of around $2.50 x 35635/1E6 = $0.09 total
actual bill was $0.09
US$0.0059/query
15 q, roughly 2k t/q, around 35635 t
expect bill of around $2.50 x 35635/1E6 = $0.09 total
actual bill was $0.09
US$0.0059/query
OpenAI gpt-3.5-turbo $0.50/Mt input, $1.50/1Mt output (cheapest old model)
7 q, 16263 t so 2323 t/q
expect bill of around 16263 x $0.60/1E6 = $0.0093 total
US$0.0014/query
7 q, 16263 t so 2323 t/q
expect bill of around 16263 x $0.60/1E6 = $0.0093 total
US$0.0014/query
OpenAI gpt-4o-mini $0.15/Mt input, $0.6/1Mt output
14 q, 31982 t, so 2284 t/q
expect bill of around 31982 x 0.16/1E6 = $0.0051 total
US$0.00037/query (estimated, the actual billing was too small to see on their graphic)
14 q, 31982 t, so 2284 t/q
expect bill of around 31982 x 0.16/1E6 = $0.0051 total
US$0.00037/query (estimated, the actual billing was too small to see on their graphic)
Compare with:
JohnsonBots Private LLM using meta-llama/llama-3-8b-instruct
L$0.1/q, $1 USD = 250 L$ so US$/q = US$ 0.1/250 /q
US$0.00040/query
L$0.1/q, $1 USD = 250 L$ so US$/q = US$ 0.1/250 /q
US$0.00040/query
So - gpt-4o-mini is comparable in price to Private LLM using meta-llama/llama-3-8b-instruct,
but 3.5-turbo and 4o are appreciably more expensive ---
--- and they are really heavily filtered now!
I tried out a bit of risque stuff and it is really hard to get OpenAI to be naughty! I was shocked!
If you want the Private LLM ConvoBot: https://marketplace.secondlife.com/p/Uncensored-Private-LLM-Generative-AI-ConvoBot-no-logging-or-filtering-no-CC-or-API-keys-pay-L-as-you-go/25570074 only L$49. No filtering, no underpaid laborers in Nigeria reading your prompts!
Details of the testing. Tokens are mostly input, output is around 2% of input for the Uma type sexbot.
Results are similar for ConvoBot or Sub Controller in most LLM setups in those systems, approximately 2200 tokens per query, depends on biographical information and conv history sent up with prompt.

Comments
Post a Comment