Gpt-realtime-1.5 text output_modality - drops words and replaces with "analysis"

API Bugs
1

I recently upgraded to gpt-realtime-1.5 for my “text” ouput_modality use cases after this previously logged bug (where control characters were leaking into text output) was reported to be fixed. The original bug does indeed seem to be fixed, but now I am seeing instances where chunks of words are replaced with the word “analysis” in the text output.

For example, in a recent session, the second response returned to me was “analysis to schedule you for a 30-minute meeting. Do you have a specific date and time in mind, or would you like me to look at our availability over the next week?”

The session ID where this occurred was sess_DcX9a7HyP09GDljdpH4vD

I have heard the word “analysis” happening fairly frequently when looking through examples. I would guess maybe 1 in 50 to 1 in 100 text responses has this issue. I have never had this issue in output_modality = “audio”, and I never had the issue on the previous gpt-realtime model.

2

We saw some new instances of this. It seems the “text” output_modality is still somewhat unusable with tool calls in gpt-realtime-1.5

In session sess_Dej6WW6AuZJTWynl1PllH, the second text response from the model was:
“assistant to=functions.search_calendar_availability deputysummary code”

In session sess_Dej8rtT5NMS1l3yWoy6tT, the second text response from the model was:
“blank”

3

I get this on almost a daily basis with text output modality. Don’t have exact stats but probably 1in a 150 ish tool calls

4

Have you figured out a nice way to work around it? Looking back at my previous posts, it occurred on the second response of the session each time. Is that the pattern you see?