This strategy raised the bar for the AI industry and competitors were forced to respond. Within days, Elon Musk’s xAI countered with Grok4, its most advanced model, now also free. Elsewhere, OpenAI effectively changed the pricing norms of frontier AI.

However, GPT‑5’s rollout was not frictionless. Despite GPT‑5’s clear advancements – less hallucinations, sharper coding skills, stronger multimodal reasoning, and improved factual precision – the decision to retire legacy models like GPT‑4o triggered swift user backlash.

Loyal users missed GPT-4o’s familiar tone, warmth, and distinct personality. GPT-5 felt colder to them – no doubt smarter, but less human. Sensing the discontent, OpenAI quickly reversed course and restored GPT‑4o for paying users. The company pledged to infuse future versions with more of the emotional nuance that made GPT-4o beloved in the first place.

This has made a very interesting case to rethink the human-AI dynamics of trust. One that could unfold across a dynamic interplay of three foundational vectors:

Competence (what it knows) × Calibration (what it admits) × Connection (how it relates)

GPT-5 moves the needle on Competence and Calibration; GPT-4o dominates Connection. At the current stage of the generative AI boom, the chatbot is still the primary interface of AI. Therefore, users reward the system that feels most reliably helpful, not just most objectively correct. That is why people prefer 4o even while acknowledging 5’s lower hallucination rate: affective trust (rapport, tone, “gets me-ness”) often decides repeat usage once baseline correctness is met.

Now, let us unpack this further.

1)    Factual Reliability (Evidenced Correctness): Competence

• What it is: Accuracy on verifiable facts and the model’s propensity to back claims with evidence or consistent reasoning.
• Why it matters: This is the floor of trust. If the floor collapses, no amount of charm saves the experience.
• GPT-5 vs GPT-4o: GPT-5 leads as it has a lower hallucination rate and more accurate long-context grounding, which raises baseline competence. GPT-4o is strong, but comparatively more prone to “confident flourishes” under ambiguity.

HealthBench is a Critical Benchmark (in Healthcare) While Examining a Model’s Hallucination Rate

Source: OpenAI

2)    Calibration

• What it is: Calibration refers to how well a model’s expressed confidence in its output matches its actual correctness. A well-calibrated model “knows what it knows” and, just as crucially, knows what it does not.
• Why it matters: Overconfidence erodes trust faster than occasional abstention.
• GPT-5 vs GPT-4o: GPT-5 wins as it is designed for factuality, correctness, and long-context reliability, which improves confidence-to-correctness alignment as a key calibration signal. GPT-5 also abstains more often or provides constrained reasoning when unsure (as observed in internal OpenAI system behavior and user feedback), indicating tighter calibration.

3)    Consistency: Connection

• What it is: Giving answers in the same default style and tone.
• Why it matters: Users form trust from repeatability and consistency.
• GPT-5 vs GPT-4o: GPT-4o outperforms. GPT-5 is not a single monolithic model – it is a router-driven unified system. Behind the scenes, a lightweight router evaluates each user prompt and dynamically routes it to one of several sub-models based on complexity and context. This architecture offers flexibility and efficiency, but it also introduces a fundamental tradeoff – you are not always talking to the same GPT-5.

That is the heart of the criticism. The experience feels inconsistent because it is inconsistent. One query might hit GPT-5’s most capable backend, while another gets served by GPT-4o or even a lighter, older variant – resulting in noticeable shifts in tone, reasoning depth, and reliability. The illusion of a unified assistant breaks down, undermining user trust in what should feel like a coherent personality.

OpenAI has acknowledged this gap and says it plans to unify the router and sub-models into a single, end-to-end architecture that will improve such consistency. But that has not happened yet.

4) Relational Warmth & Persona Fit (Affective Trust): Connection

• What it is: Nuance, tone, humor, empathy, and the sense of being “seen.”
• Why it matters: Once a model clears the accuracy bar, affinity drives stickiness and user trust.
• GPT-5 vs GPT-4o: GPT-4o leads. Users report feeling guided, not merely answered, which has increased the user attachment with the AI (and to what extent it is healthy is another topic to debate). In contrast, GPT-5, while clearly more capable in logic and reasoning, tends to come off as more formal and detached, especially in everyday interactions as a chatbot – not great with building affective trust.

5) Personalization & Memory: Connection

• What it is: Retaining preferences and domain context. Continuity of understanding and evolving goals.
• Why it matters: Continuity turns a tool into a “partner” or “friend”.
• GPT-5 vs GPT-4o: GPT‑5 may be the future, but the transition has been messy. While its superior capabilities may position it as the long-term winner, the rollout disrupted something more fragile: continuity. Many GPT‑4o users reported losing parts of their chat history, contextual memory and preferred personalized responses when GPT‑5 replaced it – effectively wiping away some shared context that had built up over time. That loss does not just break functionality; it breaks trust.

For users who saw GPT‑4o as a ‘partner/friend’ rather than a tool, it felt like starting over. And in a space where memory is relationship, that reset comes at a cost.