GPT-4 is a large language model developed by OpenAI, released on March 14, 2023, as the fourth in its Generative pre-trained transformer series. Unlike its predecessors, GPT-4 is a multimodal model that accepts both image and text inputs, allowing it to process visual information alongside written language. This capability represented a significant milestone in Artificial intelligence research, expanding the potential applications of large language models beyond text-only interactions. GPT-4 was preceded by GPT-3.5 and followed by GPT-5, and it was integrated into ChatGPT and later the OpenAI API, with a version known as GPT-4V specifically designed for image processing.
OpenAI has not publicly disclosed technical details about GPT-4, such as the precise model size, architecture, or training hardware. The technical report explicitly refrained from specifying these parameters, citing competitive and safety considerations. This lack of transparency contrasted with earlier models like GPT-3, for which more information was available. Despite this, GPT-4's capabilities were demonstrated through various benchmarks and real-world applications, including standardized tests and medical question answering.
Background
OpenAI introduced the first GPT model (GPT-1) in 2018, based on the Transformer (architecture) architecture, and trained on a large corpus of books. The following year, GPT-2 was released, a larger model capable of generating coherent text. In 2020, GPT-3 emerged with over 100 times as many parameters as GPT-2, further advancing the field. GPT-3 was refined into GPT-3.5, which powered the Chatbot product ChatGPT, launched in November 2022. The success of ChatGPT highlighted the potential of large language models, setting the stage for GPT-4's development.
Capabilities
GPT-4's default version featured an 8K context window, with a special API version supporting up to 32K tokens. As a multimodal model, it could process images and text, enabling users to upload photos and receive suggestions or answers based on the visual content. This capability was a departure from text-only predecessors and opened new avenues for interaction.
GPT-4 could also be prompted to interact with external interfaces. For example, the model could be instructed to enclose a query within <search></search> tags to perform a web search, with results inserted into the prompt to inform responses. This allowed GPT-4 to perform tasks beyond text prediction, such as using APIs, generating images, and accessing webpages.
A 2023 article in Nature noted that programmers found GPT-4 useful for coding assistance, despite its propensity for error. It could help find errors in existing code and suggest performance optimizations. A biophysicist reported that porting a program from MATLAB to Python, which previously took days, was reduced to "an hour or so" with GPT-4's help. On a test of 89 security scenarios, GPT-4 produced code vulnerable to SQL injection attacks 5% of the time, a marked improvement over GitHub Copilot from 2021, which had a 40% vulnerability rate.
In November 2023, OpenAI announced GPT-4 Turbo and GPT-4 Turbo with Vision, featuring a 128K context window and significantly lower pricing.
Aptitude on standardized tests
Studies have debated the reliability of using chatbots for standardized tests. In the Torrance Tests of Creative Thinking, GPT-4 scored within the top 1% for originality and fluency, with flexibility scores ranging from the 93rd to the 99th percentile.
Medical applications
Researchers from Microsoft tested GPT-4 on medical problems and found that, without specialized prompt crafting, it exceeded the passing score on the USMLE by over 20 points, outperforming earlier general-purpose models like GPT-3.5 and models fine-tuned on medical knowledge such as Med-PaLM. Despite this strong performance, the report warned of "significant risks" in using large language models for medical applications, including potential inaccurate recommendations and hallucinated factual errors. In April 2023, Microsoft and Epic Systems announced plans to provide healthcare providers with GPT-4-powered systems for responding to patient questions and analyzing medical records.
GPT-4o
On May 13, 2024, OpenAI introduced GPT-4o ("o" for "omni"), a successor to GPT-4 that processes and generates outputs across text, audio, and image modalities in real time. GPT-4o exhibited rapid response times comparable to human conversation, improved performance on non-English languages, and enhanced vision and audio understanding. It was also available to free-tier users, unlike GPT-4.
Limitations
Like its predecessors, GPT-4 has been known to hallucinate, producing outputs that may include information not in its training data or that contradict the user's prompt. This limitation poses challenges for applications requiring high reliability.
GPT-4 also lacks transparency in its decision-making processes. While it can provide explanations for its decisions, these explanations are formed post-hoc, and it is impossible to verify whether they reflect the actual process. In many cases, when asked to explain its logic, GPT-4 gives explanations that directly contradict its previous statements.
In 2023, researchers tested GPT-4 against ConceptARC, a benchmark designed to measure abstract reasoning, and found it scored below 33% on all categories, while specialized models scored 60% on most, and humans scored at least 91% on all. Sam Bowman, not involved in the research, suggested the results may not indicate a lack of abstract reasoning, as the test is visual while GPT-4 is a language model.
Bias
GPT-4 was trained in two stages. First, it was trained on large internet text datasets to predict the next token. Second, human reviewers used reinforcement learning from human feedback to fine-tune the system, training it to refuse prompts that go against OpenAI's definition of harmful behavior, such as illegal activities, self-harm advice, or graphic content requests. Microsoft researchers suggested GPT-4 may exhibit cognitive biases such as confirmation bias, anchoring, and base-rate neglect.
Training
OpenAI did not release technical details of GPT-4's training. The technical report refrained from specifying model size, architecture, or hardware used. It described training using supervised learning on a large dataset, followed by reinforcement learning with human and AI feedback, but omitted details on dataset construction, computing power, and hyperparameters. The report cited "the competitive landscape and the safety implications of large-scale models" as reasons for this decision.
Sam Altman stated that the cost of training GPT-4 was more than $100 million. This substantial investment reflected the scale of the model and the resources required for its development.
Reception and Impact
The release of GPT-4 generated significant interest in the AI community and beyond, highlighting the rapid progress in generative AI. Its multimodal capabilities set a new standard, influencing subsequent models like GPT-4o and Google's Gemini. The integration of GPT-4 into Microsoft's Bing Chat in February 2023 demonstrated practical applications in search and conversational AI.
GPT-4's performance on standardized tests and professional exams, such as the USMLE, sparked discussions about the potential of AI in education and healthcare. However, concerns about bias, hallucination, and transparency persisted, leading to calls for more responsible AI development.
Legacy
GPT-4 was released in ChatGPT in March 2023 and remained available until 2025, when it was removed. It was also removed from the OpenAI API on October 23, 2026. Its successor, GPT-5, continued the evolution of the GPT series. GPT-4's introduction of multimodal input processing marked a pivotal moment in AI, demonstrating that large language models could extend beyond text to understand and reason about visual information, paving the way for more versatile and integrated AI systems.