GPT-4 (Generative Pre-trained Transformer 4) is a large language model developed by OpenAI, released on March 14, 2023. It is the fourth in OpenAI's GPT foundation model series, succeeding GPT-3.5 and preceding GPT-5. Unlike its predecessors, GPT-4 is multimodal, capable of processing both text and image inputs. OpenAI has not disclosed technical details such as model size, architecture, or training hardware, citing competitive and safety concerns. The model was integrated into Microsoft's Bing Chat in February 2023 and became available in ChatGPT in March 2023, with removal from ChatGPT in 2025. GPT-4 remains accessible via OpenAI's API.
Background
OpenAI introduced the first GPT model in 2018, based on the Transformer (architecture) architecture, trained on a large corpus of books. GPT-2 followed in 2019, generating coherent text at scale. In 2020, GPT-3 expanded parameters over 100-fold compared to GPT-2, and further refinements led to GPT-3.5, which powered the ChatGPT chatbot. GPT-4 built on this lineage with enhanced capabilities and multimodal input.
Capabilities
The default GPT-4 version featured an 8K context window, with an API variant supporting up to 32K tokens. Its multimodal nature allows image and text input, enabling responses to photo uploads and spoken interactions. GPT-4 can be prompted to interact with external interfaces, such as using search tags to perform web queries, integrating results into responses, and executing tasks like API calls, image generation, and webpage summarization.
A 2023 Nature article highlighted GPT-4's utility in coding, despite error propensity. A biophysicist reported reducing a MATLAB-to-Python port from days to about an hour. In security testing across 89 scenarios, GPT-4 produced SQL-injection-vulnerable code 5% of the time, versus 40% for GitHub Copilot in 2021. In November 2023, OpenAI announced GPT-4 Turbo with a 128K context window and lower pricing.
Aptitude on standardized tests
Studies have questioned the reliability of chatbot test performance. In the Torrance Tests of Creative Thinking, GPT-4 scored in the top 1% for originality and fluency, with flexibility scores from the 93rd to 99th percentile.
Medical applications
Microsoft researchers found GPT-4 exceeded the USMLE passing score by over 20 points, outperforming GPT-3.5 and medical fine-tuned models like Med-PaLM, without specialized prompt crafting. However, they warned of significant risks, including inaccurate recommendations and hallucinated facts. In April 2023, Microsoft and Epic Systems announced GPT-4-powered systems for patient queries and medical record analysis.
GPT-4o
On May 13, 2024, OpenAI introduced GPT-4o ("o" for "omni"), processing text, audio, and image modalities in real time. It offered rapid response times, improved non-English language performance, and was available to free-tier users, unlike GPT-4.
Limitations
GPT-4 can hallucinate, producing outputs not in training data or contradicting prompts. Its decision-making lacks transparency; explanations are post-hoc and often unverifiable, sometimes contradicting prior statements. In 2023, tests on the ConceptARC benchmark for abstract reasoning showed GPT-4 scored below 33% across categories, while specialized models scored 60% and humans at least 91%. Sam Bowman noted the visual nature of the test may not reflect language model reasoning.
Bias
GPT-4 training involved two stages: unsupervised pretraining on internet text, followed by reinforcement learning from human feedback (RLHF) to refuse harmful prompts. Microsoft researchers suggested GPT-4 may exhibit cognitive biases like confirmation bias, anchoring, and base-rate neglect.
Training
OpenAI's technical report omitted model size, architecture, hardware, dataset construction, compute, and hyperparameters. Training used supervised learning and reinforcement learning with human and AI feedback. Sam Altman stated training cost exceeded $100 million, though exact figures remain undisclosed.
Reception and Impact
GPT-4 set new benchmarks in AI, influencing subsequent models from competitors like Anthropic and Google DeepMind. Its multimodal capabilities expanded applications in generative AI, from creative tools to medical diagnostics. The model's limitations, including hallucination and bias, sparked debates on AI safety and transparency, prompting research into alignment and evaluation methods.
Legacy
GPT-4's release marked a milestone in large language models, demonstrating practical utility across domains. Its successor, GPT-5, continued the series, while GPT-4 remains a reference point for multimodal AI. The model's influence persists in ongoing developments in artificial intelligence and machine learning.