💥 -> -> -> Realtime API - text input length breaks the API's ability to respond ↑

API Bugs
1

MODEL = “gpt-4o-realtime-preview”
MODEL = “gpt-4o-mini-realtime-preview”
Advertised context window: 128k tokens

  • If initial_context_message (in the code below) is over about 8000 tokens, the model no longer answers follow-up questions properly. It only states, hello, I am here to help, or something similar.
  • It also starts responding in text, when I am interacting with it by audio.
  • With shorter token inputs, it works fine: I ask in audio, it properly responds in audio.

Another user @benko.csaba noticed something similar: Context limit smaller than documented

        await stream.conversation.item.create(
            item={
                "type": "message",
                "role": "system",
                "content": [ 
                    {"type": "input_text", 
                     "text": initial_context_message}
                ],
            }

:pushpin: Update: :pushpin:

  • Right after writing this post, I figured the responses will come in too slow and without a fix since OpenAI has a lawsuit and data retention to worry about.
  • Therefore, I tried this:
    • MODEL = “gpt-4o-realtime-preview-2025-06-03”
      • This actually fixed it. However, I do not have the pockets for it. I need the mini. Thank you.