Feedback: GPT-5 vs GPT-5-Chat-Latest for Conversational AI

I’ve been testing gpt-5-chat-latest vs gpt-5 in the API for our AI mental health companion app — www.onsenapp.com.

Key Observations

  • gpt-5 is not ideal for conversational chat experiences.
    Its responses tend to:

    • Lack personality

    • Be overly structured (too many bullet points/lists)

    • Feel unnatural in the flow of a continuous conversation

  • gpt-5-chat-latest performs much better for chat, with a tone and “vibe” more similar to gpt-4.1.


The Trade-off

Since gpt-5-chat-latest does not support structured outputs, our model-selection logic has become more complex:

  • For natural conversation: gpt-5-chat-latest works best

  • For structured outputs & complex instruction-following: gpt-5 works best

Previously, gpt-4.1 (and gpt-4o before that) handled both use cases equally well, so I’m disappointed that the GPT-5 family now forces one more decision point in our prompt-engineering pipeline. This feels like a step away from the “one model to rule them all” philosophy that GPT-5 was supposed to bring.


Gap in the Portfolio

There’s currently no mini equivalent of gpt-5-chat-latest.
This creates a gap:

  • gpt-5-mini is not suitable for conversational experiences

  • There’s no drop-in replacement for gpt-4.1-mini in chat-heavy scenarios


In short:
GPT-5 offers stronger specialization but weaker versatility for teams building products that rely on both natural conversation and structured output from the same model family.

5 Likes

Maybe you just need to system prompt it to get the vibe/naturalness. Try assigning a “role” and giving adjectives how to act in the prompt

We have a sophisticated system prompt and persona role instructions. Again, GPT-5 is less able of engaging in natural, empathetic conversations vs GPT-5-Chat-Latest.

1 Like

Totally agree with your assessment. It’s always a hit or miss with the GPT-5 model. From my experience, it’s either having proper tool calls and reflection, or the conversational vibe to it. Never both.

Thank god for gpt-5-chat haha