OpenAI has released GPT-5, a new AI system featuring a real-time router that dynamically selects between fast responses and deeper "thinking" modes based on task complexity. The model demonstrates state-of-the-art performance across coding, math, writing, and health benchmarks while significantly reducing hallucinations compared to previous versions.
- GPT-5 sets new records including 94.6% on AIME 2025 (math), 74.9% on SWE-bench Verified (coding), and 84.2% on MMMU (multimodal).
- The system uses a router trained on user signals to switch between standard and extended reasoning models, with a mini version handling queries after usage limits.
- GPT-5 pro offers extended reasoning capabilities, achieving 88.4% on GPQA without tools.
- Hallucination rates are reduced by approximately 45% compared to GPT-4o and 80% compared to OpenAI o3 when using thinking mode.
- The model was trained on Microsoft Azure AI supercomputers.
This release aims to provide expert-level intelligence for real-world queries, offering more useful and accurate responses for everyday tasks in writing, coding, and health advocacy.