GPT-5 Performance: Are Increased Hallucinations and Slowness Signaling a Regression?

The evolving landscape of large language models (LLMs) often brings anticipation for continuous improvement, yet recent discussions reveal a shared sentiment among users regarding a perceived regression in the performance of GPT-5. Many individuals, particularly those previously relying on models like GPT-4, report a noticeable decline in their daily interactions.

Key Performance Concerns

Users frequently highlight several critical issues impacting their experience:

• Slowness and Inefficient Reasoning: The model is often observed entering a "thinking mode" that can be excessively long or even lead to it getting stuck. This auto-assessment mechanism appears poorly tuned, defaulting to deep reasoning even for tasks that might not require it, thereby hindering efficiency.
• Prevalence of Hallucinations: A primary complaint is the dramatic increase in hallucinations, with some users estimating a five-fold rise compared to earlier models. These inaccuracies manifest in various ways, from fabricating list items not present in the prompt to inventing software functionalities, capabilities, and even command-line interface parameters. This necessitates continuous, careful monitoring and correction, even when explicit sources are provided.
• Lack of Self-Criticality and Repetitive Outputs: Even when in its supposed "thinking mode," the model frequently produces incorrect information. Users find that a direct intervention like "this is not correct, check your answer" can often prompt it to correct itself, indicating a general lack of internal validation. Furthermore, when asked for multiple solutions or an alternative to a previous answer, the model sometimes generates identical options with different explanations or merely regurgitates its initial response with stylistic modifications, offering little new value.
• Internal Contradictions: Some report receiving responses containing direct contradictions within the same sentence, further eroding trust in the model's reliability.

Understanding the Regression: Theories and Adaptations

The discussion also delves into potential explanations and implications for user interaction:

• Cost Optimization over User Experience: One prominent theory suggests that the perceived downgrade is a strategic move, where the model employs an internal "router" to select between stronger and weaker sub-models for a given query. The primary goal of this system would be to reduce operating costs, rather than solely focusing on enhancing the user experience, despite marketing efforts to present it as an improvement.
• Model "Personalities" and User Adaptation: Another perspective posits that different LLMs possess distinct "personalities." Users transitioning from one model to another, such as from GPT-4 to GPT-5, may experience a "culture shock" requiring them to adapt their prompting styles and expectations. This implies that progression in model capabilities might not always mean a simpler user interface, but rather a different one.
• The Burden of Prompt Engineering: Many users express frustration that, contrary to expectations, the need for advanced prompt engineering seems to be increasing rather than decreasing with newer models. This adds to the cognitive load of interacting with the AI, as users must continually refine their input to achieve desired results.

Navigating the Current Landscape

While the overall sentiment leans towards disappointment, the conversation offers insights into how users are attempting to manage these challenges. Explicitly correcting the model when it errs is a demonstrated technique. More broadly, acknowledging that each model iteration might demand a recalibration of interaction strategies—similar to learning a new tool—could be a key to effectively leveraging advanced AI, even as developers work to refine consistency and reduce unforeseen regressions.