Gpt-4o-mini consumes more than 20 times tokens for images than gpt-4

API
1

For any image processing, the token usage for gpt-4o-mini is atleast 20 times higher. Just trying in playground, for a small image like this, the token usage in ~400 tokens in gpt-4o . but the same image with gpt-4o-mini is more than 8000. This is very wierd.
sample1

1 Like
2

Mini uses more tokens probably so that it stays on about the same quality as GPT-4o, but billing-wise, due to Mini’s cheap price, images cost the same for both models.

1 Like
3

In my tests (around 20 images) the cost always went down when using 4o-mini, up to 40% in some images. The task was data extraction from invoices and quality-wise I didn’t find any difference, but be careful, this really depends on your specific use case.

Btw, if you are using Langchain, libraries haven’t been updated yet and they default calls to 4o, even if you explicitly use “gpt-4o-mini” as model.

4

Here is about dozen similar topics that gives totally same answer. Just saying.

5

For the same set of images, gpt-4o statistics are:

"usage": {
        "prompt_tokens": 12288,
        "completion_tokens": 1912,
        "total_tokens": 14200
    }

But gpt-4o-mini’s are:

"usage": {
        "prompt_tokens": 408064,
        "completion_tokens": 439,
        "total_tokens": 408503
    }

This is way more than expected. The final cost ended up being ~9c for gpt-4o and ~6c for gpt-4o-mini though.

The pricing web page does show that for a 150x150 image, gpt-4o-mini can be more expensive than some gpt-4o models ($0.001275 with Mini vs. $0.000638 with gpt-4o-2024-08-06) and it does use much more tokens for images (8500 vs. 255).

This is simply absurd given that gpt-4o-mini is marketed as OpenAI’s “most cost-efficient small model”. Not for vision it seems.