GPT-4

GPT-4 was released on 14 March 2023 and is the model against which the current generation is still implicitly measured. It was OpenAI's first flagship that accepted images as well as text, and the first whose evaluation was framed around human exams rather than research benchmarks: it passed a simulated Uniform Bar Examination in roughly the top 10% of test takers, where its predecessor had scored in the bottom 10%. Technically OpenAI disclosed almost nothing. The accompanying paper stated openly that it contained no details of architecture, model size, hardware, training compute, dataset construction or training method, citing competitive and safety considerations — a break with the publishing norms of the preceding decade that the rest of the industry then followed. What was published: a context window of 8,192 tokens with a 32,768-token variant, a knowledge cutoff in September 2021, and scores including 86.4% on MMLU and around 67% on HumanEval. The model was extended rather than replaced for over a year. GPT-4 Turbo, announced in November 2023, raised the context window to 128,000 tokens and moved the cutoff forward; GPT-4o followed in May 2024 with native speech and vision. GPT-4 itself was removed from ChatGPT on 30 April 2025 and now survives only as legacy API snapshots: gpt-4-0314 was shut down on 26 March 2026 and gpt-4-0613 is scheduled to follow on 23 October 2026. Its wider significance is less about capability than about timing. GPT-4 arrived three months after ChatGPT had already made the technology famous, and it converted that attention into an enterprise market — while simultaneously triggering the open letter calling for a six-month pause on systems more powerful than GPT-4, the EU AI Act's late redrafting around general-purpose models, and the safety-institute apparatus that followed.

▸Videos

No videos yet.

▸Reviews