From GPT-3 to GPT-5: Mapping their capabilities, scope, limitations, and consequences

A comprehensive comparative study traces the evolution of OpenAI's GPT models from GPT-3 through GPT-5, revealing that successive generations represent far more than incremental capability improvements. The research demonstrates a fundamental shift from simple text predictors to integrated, multimodal systems with tool access and workflow capabilities, while persistent limitations like hallucination and benchmark fragility remain largely unresolved across all versions.

This arXiv paper provides essential scholarly grounding for understanding how large language models have transformed from research artifacts into production systems. The authors challenge the common narrative that GPT improvements are merely quantitative, arguing instead that each generation represents a qualitative reformulation of what deployable AI systems are and how responsibility is distributed when they operate at scale. This distinction matters because it reframes how stakeholders—developers, enterprises, and regulators—should evaluate these technologies.

The research documents how the GPT family evolved across five dimensions: technical architecture, user interaction patterns, multimodal capabilities, deployment infrastructure, and governance frameworks. Earlier generations functioned as few-shot text predictors; later versions integrate tool access, extended context windows, and safety-tuning mechanisms that fundamentally alter their effective capabilities. This means direct model-to-model comparisons obscure the true innovations, which lie in system design rather than raw parameter scaling.

For the AI industry, this analysis highlights why enterprise adoption hinges on more than raw accuracy metrics. Organizations implementing GPT systems must account for safety mechanisms, interface design, tool integration, and governance—factors that vary significantly across versions and directly impact deployment outcomes. The persistence of core limitations like hallucination and prompt sensitivity across all generations signals that these may be inherent properties rather than engineering challenges. For investors and developers, the implication is clear: future differentiation emerges not from model size alone but from how systems are architected, evaluated, and integrated into workflows. The paper suggests that continued progress requires rethinking evaluation frameworks and responsibility structures rather than pursuing incremental capability scaling.

• →GPT evolution represents a shift from text prediction to integrated multimodal tool-oriented systems, not merely larger models
• →Core limitations including hallucination, prompt sensitivity, and benchmark fragility persist across all GPT generations unchanged
• →Effective system capability now depends on routing, tool access, safety tuning, and interface design—not model capability alone
• →Public transparency about architecture and training remains incomplete despite rapid deployment of increasingly powerful systems
• →Future progress requires rethinking evaluation frameworks and responsibility location for frontier AI systems at scale