Cost experience with OpenAI

[dennisvandehoef]

Heey,

Feel free to close this right away.

I implemented paperless-ai today.
After a week trying to estimate the costs.

This issue is there to help all people like me, hoping to find out the costs before starting.

I have used up to 25 documents, to fine tune the prompt. Removing all tags / corospondents / document types in paperless for atleast 10 times.
Then I told it to index all my 156 documents (fresh paperless).

Currently my open-ai bill is at a stunning (not) 0,10 us dollar.

Sharing this, as I found it hard to estimate the cost of this project.'
Hoping that this issue helps atleast 1 person in the future.

Best,
Dennis

[clusterzx]

What model did you use? How big are your files?

I use gpt-4o-mini and have used over 7.6 million tokens.

I am arround 2,11$ right now.

[Daniel15]

Does this count both input and output tokens, or only input?

[dennisvandehoef]

You are right, I forgot to add some important information

I also used gpt-4o-mini and used almost 600k tokens, all documents together have 565k Charakters

[clusterzx]

I think the 0.10$ are very reasonable in that case, right? If you think about it that 1$ could last you for the next years if you just do your regular documents.

[dennisvandehoef]

Yea that definitely is reasonable.

[Magneticdud]

Especially because openai pockets the remainder of the $5 credit 12 months after purchase because "it expires".

For a personal use it's very hard to use that credit within a year

[rknightion]

I asked the AI to analyse the costs for me :). using 4o-mini I am amazed by how cheap it is! https://chatgpt.com/share/67a6654b-0970-8001-848f-96168b33bcde

[Daniel15]

@Dulanic It works well! I don't have any complaints. I've scanned maybe 50 documents (including some fairly large ones) and it's only cost one cent.

On average, the API call takes 0.9 seconds, which is pretty fast. I set paperless-ai up as a webhook (instructions are in the wiki) so it immediately runs as soon as a new document is added. After I scan a document, by the time I load/refresh paperless-ngx, paperless-ai has already processed it.

[princejosuah]

For what it's worth, Google's Gemini 2.0 Flash is even cheaper, and I trust Google's privacy policy (on the paid version, not the free one) more than OpenAI's. Also, Google is profitable, so their API is more likely to survive in the long run.

I'm a bit at a loss... How did you test with Gemini? I can't see it among the options

[clusterzx]

For what it's worth, Google's Gemini 2.0 Flash is even cheaper, and I trust Google's privacy policy (on the paid version, not the free one) more than OpenAI's. Also, Google is profitable, so their API is more likely to survive in the long run.

I'm a bit at a loss... How did you test with Gemini? I can't see it among the options

Custom API setting

[Daniel15]

For what it's worth, Google's Gemini 2.0 Flash is even cheaper, and I trust Google's privacy policy (on the paid version, not the free one) more than OpenAI's. Also, Google is profitable, so their API is more likely to survive in the long run.

I'm a bit at a loss... How did you test with Gemini? I can't see it among the options

@princejosuah

AI Provider: Custom
Base URL: https://generativelanguage.googleapis.com/v1beta/openai/
API Key: Create one at https://aistudio.google.com/app/apikey
Model: gemini-2.0-flash

Make sure you eventually set up billing, as Google can retain all data used with the free model and use it for training, whereas they have a good privacy policy for the paid version (since it's used by businesses). The free version is okay for testing purposes as long as you're OK with Google keeping the documents.

[princejosuah]

For what it's worth, Google's Gemini 2.0 Flash is even cheaper, and I trust Google's privacy policy (on the paid version, not the free one) more than OpenAI's. Also, Google is profitable, so their API is more likely to survive in the long run.

I'm a bit at a loss... How did you test with Gemini? I can't see it among the options

@princejosuah

AI Provider: Custom
Base URL: https://generativelanguage.googleapis.com/v1beta/openai/
API Key: Create one at https://aistudio.google.com/app/apikey
Model: gemini-2.0-flash

Make sure you eventually set up billing, as Google can retain all data used with the free model and use it for training. The free version is okay for testing purposes as long as you're OK with Google keeping the documents.

Thank you so much for the detailed response, I'll be testing this ASAP!