GPT-4 is a large-scale multimodal Large language model developed by OpenAI and released on March 14, 2023, as the successor to GPT-3 and the underlying model powering an upgraded version of ChatGPT. Unlike its predecessors, GPT-4 accepts both text and image inputs, though its output remained text-only in its original release, and OpenAI reported substantially improved performance across a wide range of professional and academic benchmarks, including passing a simulated bar exam in a percentile OpenAI compared favorably to human test-takers.
Secrecy and technical disclosure
GPT-4's release marked a notable shift in OpenAI's publication practices. Unlike earlier GPT papers, which detailed architecture and training data, OpenAI's GPT-4 technical report disclosed no information about model size, architecture details, training compute, or dataset composition, citing competitive and safety considerations. This departure from OpenAI's earlier norm of openness drew criticism from parts of the research community, who argued it undermined the scientific reproducibility that had characterized the field, and became a frequently cited example in broader debates over transparency among developers of frontier models.
Sparks of AGI
Shortly after release, a group of Microsoft researchers published "Sparks of Artificial General Intelligence," a paper examining GPT-4's capabilities across coding, mathematics, medicine, law, and other domains, arguing the model displayed early, general-purpose problem-solving abilities that went beyond narrow pattern matching, and could reasonably be discussed in relation to Artificial general intelligence. The paper was influential in shifting public and research discourse, though it also drew methodological criticism for relying on anecdotal, non-systematic evaluation rather than rigorous benchmarking.
Training and safety
GPT-4 was trained using large-scale pretraining followed by RLHF to align its outputs with human preferences, alongside extensive red-teaming conducted with external researchers before release, an application of practices associated with Red teaming (AI) and Guardrails (AI) work across the industry. OpenAI later released GPT-4 Turbo and other iterative updates that expanded context length and reduced cost, and offered GPT-4 as the default model behind paid ChatGPT tiers, contributing to sustained subscription revenue growth.
Impact
GPT-4's release intensified competitive pressure across the AI industry, prompting accelerated releases from Google DeepMind, Anthropic, and other labs, and is frequently cited as the point at which frontier language model development shifted decisively from academic-adjacent research culture toward large-scale, capital-intensive commercial competition, a period sometimes referred to as part of the broader AI boom (2022-present). OpenAI later released OpenAI o1 in 2024 as a distinct reasoning-focused successor line built on top of GPT-4-class base models.