I encountered an issue with the current model selection for the OpenAI service. It defaults to gpt-4-turbo-preview as the preferred multimodal model, but this model fails to generate image alt text, returning the error:

“Invalid content type. image_url is only supported by certain models.”

This error persists even though a valid base64-encoded image is sent in the request.

To resolve this, I’ve updated the model selection to use only gpt-4o models, which are compatible and successfully handle the image input for generating alt text.

I also excluded realtime and audio models. Realtime models are incompatiable with the v1/chat/completions endpoint (they require v1/completions endpoint). Audio models don't accept image inputs.

Let me know if this fits or if you need any more tweaks!