GPT-4: The Reference Model, and What It Withheld

GPT-4, released in March 2023, was OpenAI's first large multimodal model, taking image and text input and producing text. It posted human-level results on a range of professional and academic exams and was paired with a documented red-team process involving more than fifty outside experts across alignment, cybersecurity, and biorisk. Two things make the release worth remembering beyond the capability jump. First, the technical report explicitly declined to publish architecture, model size, training data, or training method, citing competitive and safety concerns. That was a turning point in how openly frontier work gets described, and most labs followed. Second, the accompanying system card was unusually detailed about limitations: hallucination, a fixed context window, no learning from experience, and concrete analysis of bias, disinformation, and proliferation risk. OpenAI was clear that the model was confidently wrong in predictable ways.

Why it matters

GPT-4 became the bar that nearly every later model, including the fully open ones, measures itself against, so understanding what it actually claimed is useful context. The reporting precedent matters too: this is when "frontier model" started to mean you do not get to see how it was built.