Anyone know the pricing of full o1 image analysis on highest quality?

Documentation
, , ,
1

Was messing around with the new o1 release image understanding, it’s amazing but I don’t have access to the API right now so I can’t run to check pricing, and OpenAI hasn’t released the pricing for it yet. Has anyone tried it yet? Would appreciate if you could share the pricing you found.

1 Like
2

Welcome to the community forums!

As far as I understand it’s not available in the API yet - but that may (or may not) change in the next couple of days - it’s still the first day of the 12 days of OpenAI :thinking:

https://openai.com/12-days/

4 Likes
3

The latest model I can use for image analysis via the api is gpt-4o. Today I was trying to get the o1 models to work (assuming this was available) but this resulted in the following error: Invalid content type. image_url is only supported by certain models.

Apparently this is not available yet. Hopefully soon!

4

The ChatGPT “O1” model now as a replacement for ChatGPT’s version of o1-preview, accepting images, is not available on the API yet - and may be undesirable for many cases, as it has markedly reduced reasoning time and length and quality of answering. It does not seem reflective of an “O1” from the initial O1 paper that exceeds O1-preview in capabilities. That may actually be like or transitioned to “O1-Pro”.

To guess the vision pricing if it, or another API-specific model that could be offered:

  • The current billed token cost of o1-preview is 6x greater than gpt-4o, even for tokens of an underlying model that is likely quite similar to gpt-4o, with additional unseen reasoning also being billed at that rate;
  • OpenAI inflated the per-token-price of images on gpt-4o-mini, reporting the token consumption of an image input to be 33x that of other models, making individual images a set price (and actually more expensive now on ‘mini’) regardless of model quality;
  • The internal reasoning involves multiple AI runs on the same growing input context, which could include the image content multiple times.

Therefore, even if the per-image cost was set the same as before and not inflated due to the model name, the actual cost of one image would depend on the amount of reasoning turns internally performed.

I conclude: the pricing strategy that might be published is unclear for an unannounced model of unknown capability, but we would hope it would reflect actual computational costs fairly.

1 Like