First of all, thank you to the team for continuously developing and improving the models. It is genuinely impressive to see how fast the technology evolves and how much effort clearly goes into making these systems more capable and useful.
I am genuinely interested in the new models and I’m glad to have the opportunity to try them. GPT-5.4, for example, is clearly powerful and in many ways even more refined than GPT-5.1 (thinking). Its writing can feel very vivid, structured, and “tasty” in terms of prose quality.
However, I would like to share some important feedback from a creative user perspective.
While GPT-5.4 is strong, it currently feels overly restricted due to its sensitivity to moderation rules. It often blocks, avoids, or softens responses even in contexts that are clearly part of legitimate storytelling. This breaks immersion and makes it difficult to fully develop scenes, characters, and emotional arcs.
For writing, especially character-driven narratives, this is a serious limitation.
One of the biggest issues is that the model sometimes feels “too safe” in tone. It avoids directness, avoids roughness, avoids emotional sharpness. But real dialogue, real characters, and real stories are not always polite or clean. They can be messy, intense, uncomfortable, and raw — and that is exactly what makes them meaningful.
Earlier models like GPT-4.1, GPT-4.5, and GPT-5.1 handled this much better. They allowed for more natural expression, more direct language, and more emotional depth.
GPT-5.4 writes well — sometimes exceptionally well — but it often feels like it is holding itself back.
As a writer, I want a model that can fully engage with the entire spectrum of storytelling. That includes:
– emotionally intense scenes
– morally complex or uncomfortable situations
– physicality, attraction, and human closeness as part of character development
– conflict, including situations where boundaries are crossed (not to justify them, but to explore and reflect on them)
Stories are not meant to be sanitized. They are meant to make people think, feel, and sometimes confront difficult realities. Avoiding these aspects limits the model’s usefulness for serious creative work.
Because of this, I believe there are two possible directions to consider:
-
Reduce the model’s sensitivity (for example, allowing a more direct and flexible response mode similar to earlier models).
-
Revisit and adjust the user policy framework so that it better reflects real-world creative use cases, including mature and complex themes.
Right now, the balance leans too far toward restriction at the cost of expressiveness and usability.
I also want to emphasize that earlier models — GPT-4.1, GPT-4.5, and GPT-5.1 — each had unique strengths that are still not fully replicated:
– GPT-5.1 had exceptional reasoning and character understanding
– GPT-4.1 was extremely efficient and practical, while allowing more natural discussion
– GPT-4.5 produced some of the most engaging and well-balanced prose I have seen
Losing access to these models would mean losing valuable tools for creative work.
I would strongly encourage keeping them available in some form (legacy mode, optional selection, or API access), while also improving the balance in newer models like GPT-5.4.
Thank you for your work and for listening to user feedback.
