OpenAI announced GPT-4.5, trained on Azure AI supercomputers, as a research preview advancing the "unsupervised learning" scaling paradigm via increased compute and data. The model shows benchmark improvements over GPT-4o on science (71.4% vs 53.6% on GPQA), math (36.7% vs 9.3% on AIME), and coding tasks, with claimed reductions in hallucination and improvements in semantic understanding, nuance, and "emotional intelligence" in conversation. GPT-4.5 is positioned as a general-purpose foundation model separate from the reasoning models (o1/o3-mini), with availability first to Pro subscribers and then to Plus, Team, Enterprise, and Edu users; API access is in preview but explicitly not replacing GPT-4o and uncertain for long-term deployment. The release directly exposes the two-axis strategy: unsupervised learning (incremental, broad capability) vs reasoning (chain-of-thought, hard STEM problems), and underscores that frontier model generation cycles are now tied to both compute scaling and architectural/training innovation.
524 other tracked developments touching the same claims.