We’ve thoroughly tested these images and confirmed that they are accessible while OpenAI is processing the requests. Additionally, testing with static, publicly hosted images yields the same results (i.e., errors).
The issue appears more frequently when we pass a higher number of image links—typically above 50. Due to the size and quantity of images, passing them as base64 is not a viable alternative for us.
This seems to be a known issue (see similar posts in community), but so far, we haven’t identified any solution, and it’s becoming critical for our business.
As it is not a reproducible error every time that leads me to think it is related to availability. As OpenAI does not publish in real time the load and stats the best one can do is look at an unofficial posting of such.
While I do not look at the unofficial status daily I do know that they change in small fluctuations and large fluctuations so it may take a few weeks before you will see a large fluctuation for better or worse.
Also when models become more available for public use, as in moving out of beta, preview, etc. this tends to increase load. Another speculation is that Orion model is being load tested before release and this loading is a pattern that some of us have observed in the past.
Thanks for your answer.
Given how often it occurs (sometimes > 10% of calls) it seems unlikely to me that it is related to availability though - or I’m maybe missing your point?
