The article titled “Why GPT-5 says GPT-4…” explores the evolving relationship between successive generations of large language models and challenges the assumption that each new release fully replaces its predecessor.
***
Shift from creativity to precision
One of the central arguments highlighted in the analysis is that GPT-5 represents a shift in design philosophy. While earlier models like GPT-4 were widely praised for their creativity, conversational tone and adaptability, newer iterations prioritise structured reasoning, efficiency and accuracy.
Industry observations support this trend. Reports suggest that GPT-5-series models are more optimised for analytical tasks, including coding, data processing and complex reasoning workflows, rather than open-ended creative interaction.
At the same time, some users and analysts argue that this shift has come at a cost. Feedback from both media reports and community discussions indicates that GPT-5 can feel more rigid and less expressive compared to GPT-4-era models, particularly in writing and brainstorming contexts.
Architectural evolution changes behaviour
Another key factor behind the perceived differences lies in how newer models are built. GPT-5 operates as a multi-layered system that dynamically routes queries between faster and deeper reasoning subsystems depending on task complexity.
This architecture improves efficiency and reduces unnecessary computation, but it can also produce responses that feel shorter or more constrained. Analysts note that such optimisation may explain why some users perceive GPT-5 as less nuanced in certain scenarios.
Performance gains in enterprise use cases
Despite criticism, the article emphasises that GPT-5 and its newer variants bring measurable improvements in professional environments. These include reduced error rates, lower token usage and better performance in structured workflows.
For example, newer models have been reported to use significantly fewer tokens while maintaining or improving output quality, lowering operational costs in large-scale deployments.
Additionally, GPT-5-series models are increasingly positioned as tools for enterprise automation rather than general-purpose experimentation, reflecting a broader shift in how AI products are monetised and deployed.
User backlash highlights trade-offs
The transition from GPT-4 to GPT-5 has not been without controversy. Following early releases, some users expressed dissatisfaction with the newer model’s tone, flexibility and perceived reliability.
Media coverage indicates that complaints focused on shorter responses, reduced personality and limited creative range, prompting companies to reintroduce older models alongside newer ones to maintain user choice.
This backlash underscores a broader reality: improvements in technical capability do not always align with user expectations.
Why GPT-4 still remains relevant
The LaoZhang analysis ultimately argues that GPT-4 remains valuable because it excels in areas where GPT-5 is less dominant. These include:
Creative writing and storytelling
Conversational engagement and tone variation
Flexible brainstorming tasks
In contrast, GPT-5 is better suited for:
Logical reasoning and structured outputs
Coding and technical problem solving
Enterprise-level automation
This divergence suggests that AI development is no longer strictly linear, where each version replaces the last, but instead increasingly specialised.
A broader industry trend
The debate reflects a wider shift across the AI sector. As models mature, progress is becoming more incremental and use-case driven rather than revolutionary.
Some analysts argue that the industry is entering a phase of optimisation, where improvements focus on efficiency, reliability and cost rather than dramatic leaps in capability.
At the same time, research continues to highlight that even advanced models have limitations, particularly in reasoning consistency and complex problem solving, reinforcing the need for ongoing development.
Conclusion
The claim that GPT-5 “replaces” GPT-4 is increasingly being challenged. Instead, emerging evidence suggests a more nuanced reality: newer models are not universally better, but differently optimised.
As AI systems continue to evolve, the choice between models may depend less on version number and more on the specific task at hand.
