Hi all, we have a production agent running on `gpt...
# agents
b
Hi all, we have a production agent running on
gpt-4o-realtime-preview (2025-06-03)
which seems to massively outperform the new GA
gpt-realtime
. intelligence, speed, accuracy, tool calls, accent consistency etc… In my experience the new GA seems like a huge backstep? Curious about other peoples experience with these models. I mention this because Azure seem to be canning off 4o-realtime-preview in favour of the GA (understandable) so I’m not sure if its my wonky prompting and/or config or if other people are experiencing the same?
a
We have similar experience.. However they also mentioned they changed the way they follow prompts. So I figured its a prompt thing. Out of curiosity - how do you prompt it? In my experience giving it too much information confuses it - at least for roleplaying scenarios
b
the prompt has actually grown quite big to be fair, the agent handles menu/food ordering so accuracy and flow are super important. The prompt follows the realtime guidelines format quite closely. I went down the path of editing the prompt for the new gpt-realtime GA but its still lacking imo. Do you have a source for the mentioned changing of following prompts?
a
https://openai.com/index/introducing-gpt-realtime/ "It’s better at interpreting system messages and developer prompts" - not exactly sure what that means. Dont think openAI even knows 😄
b
ahh yeah I’ve seen this… I’m glad my SIP agent can now upload photos... 😒 - I do feel better knowing its not just me then ha. Cheers!