GPT-4 is a large language model developed by OpenAI, the fourth in its GPT foundation model series. Released on March 14, 2023, it succeeded GPT-3.5 and was later followed by GPT-5. Unlike its predecessors, GPT-4 is multimodal, meaning it can process both text and images as input. OpenAI has not disclosed technical details such as the model's size, architecture, or training hardware, citing competitive and safety concerns. An early version was integrated into Microsoft's Bing Chat in February 2023, and GPT-4 was available in ChatGPT from March 2023 until its removal in 2025; it remains accessible via OpenAI's API.
Background
OpenAI introduced the first GPT model in 2018, based on the Transformer (architecture) architecture and trained on a large corpus of books. GPT-2 followed in 2019, generating coherent text at scale. In 2020, GPT-3 was released with over 100 times more parameters than GPT-2, and its refinement led to GPT-3.5, which powered the ChatGPT chatbot. GPT-4 built on this lineage, adding multimodal capabilities and improved performance across many benchmarks.
Capabilities
The default GPT-4 version had an 8K context window, with a special API variant supporting up to 32K tokens. As a multimodal model, it can accept image inputs, enabling users to upload photos for analysis or suggestions. GPT-4 can also interact with external interfaces, such as performing web searches by enclosing queries in specific tags, allowing it to access and summarize webpages, use APIs, and generate images. A 2023 article in Nature noted that programmers found GPT-4 useful for coding assistance, such as identifying errors and suggesting optimizations, despite a tendency to make mistakes. In security tests, GPT-4 produced code vulnerable to SQL injection in 5% of scenarios, compared to 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 running chatbots through standardized tests. In the Torrance Tests of Creative Thinking, GPT-4 scored in the top 1% for originality and fluency, with flexibility scores ranging from the 93rd to 99th percentile.
Medical applications
Microsoft researchers found that GPT-4 exceeded the passing score on the USMLE by over 20 points without specialized prompting, outperforming GPT-3.5 and medical-specific models like Med-PaLM. However, they warned of significant risks, including inaccurate recommendations and hallucinations. In April 2023, Microsoft and Epic Systems announced plans to provide healthcare providers with GPT-4-powered systems for patient queries and medical record analysis.
GPT-4o
On May 13, 2024, OpenAI introduced GPT-4o ("o" for "omni"), a successor that processes and generates text, audio, and image modalities in real time. It offers response times comparable to human conversation, improved non-English language performance, and enhanced vision and audio understanding. GPT-4o was also available to free-tier users, unlike GPT-4.
Limitations
GPT-4, like its predecessors, can hallucinate, producing outputs that contradict training data or user prompts. It also lacks transparency in decision-making; explanations are formed post-hoc and may contradict prior statements. In 2023, tests using the ConceptARC benchmark showed GPT-4 scored below 33% on abstract reasoning tasks, while specialized models scored around 60% and humans at least 91%. Researchers noted this may not indicate a lack of abstract reasoning, as the test is visual and GPT-4 is a language model.
Bias
GPT-4 was trained in two stages: first, on large internet text datasets to predict next tokens; second, using reinforcement learning from human feedback to align with safety guidelines. 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, including model size, architecture, hardware, or hyperparameters. The technical report described a combination of supervised learning and reinforcement learning with human and AI feedback, but omitted specifics. Sam Altman stated the training cost exceeded $100 million. The report cited competitive landscape and safety implications as reasons for withholding information.