GPT-4

GPT-4 is a large multimodal model created by OpenAI as the fourth generation in its GPT series of generative pre-trained transformers, built on the transformer lineage that OpenAI has scaled progressively since GPT-1. According to OpenAI's own research page, it accepts image and text inputs and emits text outputs, and was announced as a milestone in scaling up deep learning after a six-month iterative alignment process informed by adversarial testing and lessons from ChatGPT. The associated technical paper is available on arXiv, and OpenAI published a system card detailing its evaluation methodology alongside the announcement.

In practical terms, GPT-4 was positioned by OpenAI as exhibiting human-level performance on a range of professional and academic benchmarks, including passing a simulated bar exam in roughly the top ten percent of test takers, a notable step up from GPT-3.5's bottom-ten-percent performance on the same assessment. The model's deep-learning infrastructure was rebuilt from the ground up together with Azure, with GPT-3.5 serving as a test run that surfaced bugs and improved theoretical foundations before GPT-4's full training. These characteristics make it well suited for complex reasoning, multimodal understanding, and benchmark-grade tasks where strong general knowledge and instruction following matter more than lightweight conversational use.