From GPT-3 to GPT-5: Mapping their capabilities, scope, limitations, and consequences
Organizations: MediaFutures, Department of Information Science and Media Studies, University of Bergen, Bergen, Norway. · Faculty of Science and Technology, Norwegian University of Life Sciences, Ås, Norway. · Faculty of Computing and Information Technology, Sohar University, Sohar, Oman. · Department of Computer Science (IDI), Norwegian University of Science and Technology, Gjøvik, Norway. ·
Abstract
We present the progress of the GPT family from GPT-3 through GPT-3.5, GPT-4, GPT-4 Turbo, GPT-4o, GPT-4.1, and the GPT-5 family. Our work is comparative rather than merely historical. We investigates how the family evolved in technical framing, user interaction, modality, deployment architecture, and governance viewpoint. The work focuses on five recurring themes: technical progression, capability changes, deployment shifts, persistent limitations, and downstream consequences. In term of research design, we consider official technical reports, system cards, API and model documentation, product announcements, release notes, and peer-reviewed secondary studies. A primary assertion is that later GPT generations should not be interpreted only as larger or more accurate language models. Instead, the family evolves from a scaled few-shot text predictor into a set of aligned, multimodal, tool-oriented, long-context, and increasingly workflow-integrated systems. This development complicates simple model-to-model comparison because product routing, tool access, safety tuning, and interface design become part of the effective system. Across generations, several limitations remain unchanged: hallucination, prompt sensitivity, benchmark fragility, uneven behavior across domains and populations, and incomplete public transparency about architecture and training. However, the family has evolved software development, educational practice, information work, interface design, and discussions of frontier-model governance. We infer that the transition from GPT-3 to GPT-5 is best understood not only as an improvement in model capability, but also as a broader reformulation of what a deployable AI system is, how it is evaluated, and where responsibility should be located when such systems are used at scale.
Figures & tables
| Model | Interaction style | Modality framing | Context window | Max output | Operational framing |
|---|---|---|---|---|---|
| GPT-3 | Prompt-completion, few-shot in context | Text in, text out | 2,048 tokens | Within context budget | Few-shot prompting; no chat-first framing; publicly presented as a large autoregressive language model. |
| GPT-3.5 Turbo | Chat-optimized dialog | Text in, text out | 16,385 tokens | 4,096 tokens | Chat Completions optimization; useful for code and non-chat tasks; publicly framed as a legacy chat model for instruction-following. |
| GPT-4 Turbo | Assistant model with long context | Text and image in, text out | 128,000 tokens | 4,096 tokens | More deployable GPT-4 variant for long-context workflows; publicly framed as a cheaper, better GPT-4 and now legacy relative to newer models. |
| GPT-4o | Real-time omni assistant | text, audio, image, video input; text, audio, image output | 128,000 tokens | 16,384 tokens | Multimodal interaction with later specialized realtime/audio variants and broader safety focus. |
| GPT-4.1 | Developer-oriented non-reasoning model | Text and image in, text out | 1,047,576 tokens | 32,768 tokens | Explicit tool calling, long context, and agent-enabling workflows; publicly framed as non-reasoning model for coding, instruction following, and large-context tasks. |
| GPT-5.x / GPT-5.4 | Routed frontier workflow system / flagship API model | Text and image in, text out in current flagship API | 1,050,000 tokens (GPT-5.4) | 128,000 tokens (GPT-5.4) | Configurable reasoning effort, tools, routed fast-vs-thinking behavior, and safe-completions; publicly framed as a frontier model for agentic, coding, and professional workflows. |
| Failure mode | Brief meaning | GPT-family implication |
|---|---|---|
| Hallucination and factual instability | Hallucination is the generation of false, invented, or unsupported claims. Factual instability is the broader tendency of factual answers to vary across prompts, runs, or contexts. | Later GPT models reduce some factual errors, but the problem is not eliminated. GPT-4 acknowledges hallucination-related limitations, while GPT-5 reports improvement without claiming full resolution. |
| Prompt sensitivity and specification fragility | Model behavior can change substantially with small differences in wording, ordering, formatting, turn structure, or tool configuration. | GPT-3 made prompt sensitivity highly visible in few-shot prompting, and later instruction-tuned models still remain sensitive to how tasks are specified and scaffolded. |
| Reliability under long reasoning chains | The model may produce fluent multi-step reasoning but still fail by dropping constraints or compounding small intermediate errors. | Longer reasoning traces and larger context windows do not guarantee correctness. |
| Benchmark contamination and comparability concerns | Benchmark contamination occurs when evaluation items, or close variants, appear in training data. Comparability concerns arise when cross-model comparisons use different prompts, tools, benchmarks, or evaluation pipelines. | Reported gains may reflect real modeling progress, but may also partly reflect benchmark fit, prompting style, tool scaffolding, or changing evaluation conditions. |
| Bias and uneven performance | Bias refers to systematic socially harmful skew, stereotypes, or exclusion. Uneven performance means model quality is not uniform across groups, languages, tasks, or modalities. | GPT-4 and GPT-4o safety materials discuss disparities and representational harms, and multimodal systems can extend these concerns through voice, image, and multilingual interaction. |
| Transparency limits and incomplete disclosure | Transparency limits are boundaries on what developers reveal publicly. Incomplete disclosure means key technical and evaluative details needed for replication or auditing are missing. | Later GPT generations are increasingly documented through partial technical reports, system cards, launch posts, and model pages, limiting transparency. |
Language Modeling
Language modeling develops systems that learn patterns in text and other sequences to understand, predict, and generate language.
See all 519 papers