Realtime sessions through the AI Gateway can't take an image. The normalized protocol only has text, audio, and function-call-output conversation items, so there's no way to hand a photo to openai/gpt-realtime-2.1 even though the underlying Realtime API accepts image input. The other problem is that sending one anyway fails silently: no error, no warning, the item simply never appears and the model answers as though nothing was sent. I tried three different shapes, including OpenAI's own native input_image shape, and all three were quietly discarded. In one case the Gateway echoed the item back with the text intact and the image stripped out, which is how I eventually worked out what was happening.

Please let me know if I'm just doing this wrong.