GPT-4.5

GPT-4.5 was released on 27 February 2025 as a research preview, and it is the clearest marker of a turning point in how frontier models are built. It was the largest and most compute-intensive model OpenAI had shipped, and it was built the old way: by scaling unsupervised pre-training rather than by teaching the model to reason for longer before answering. OpenAI presented it as a model with broader world knowledge, better instinct for what a user meant and a warmer writing style — not as a model that would win mathematics competitions. On the measure that matters most for ordinary use, it did deliver. OpenAI reported 62.5% accuracy on SimpleQA, a factual-recall test, with a hallucination rate of 37.1% — against 38.2% accuracy and 61.8% hallucinations for GPT-4o. Fewer confident inventions is a real improvement, and one that benchmarks of reasoning do not capture. The price was extraordinary: 75 dollars per million input tokens and 150 per million output, twenty-five times GPT-4o. OpenAI said openly that it was evaluating whether to keep serving the model at all, given the GPU capacity it consumed. The answer came quickly. The model was deprecated on 14 April 2025, seven weeks after launch, and removed from the API on 14 July 2025 with gpt-4.1 as the recommended replacement — a working life of four and a half months. It survived longer in ChatGPT, where paying subscribers could still select it, and its removal there on 26 August 2026 takes the last member of the GPT-4 family out of the consumer product. What it demonstrated has outlasted it: scaling pre-training alone was no longer worth the money, and the industry's compute went into reasoning instead.

▸Videos

No videos yet.

▸Reviews