X @外汇交易员

Model Performance - OpenAI releases GPT-5 series with four versions: GPT-5, GPT-5 mini, GPT-5 nano, and GPT-5 Pro (enterprise and $200/month premium) [1] - GPT-5 achieves a 749% first-try accuracy on SWE-bench Verified benchmark, surpassing o3's 691% and GPT-4o's 308% [1] - GPT-5 Pro scores 42% on Humanity's Last Exam with tools, slightly below Grok 4 Heavy's 444% [2] - OpenAI claims GPT-5 is more reliable, practical, and reduces hallucinations [2] Programming Capabilities - GPT-5 excels in complex front-end generation and large codebase debugging, creating responsive websites, apps, and games from a single prompt [1] Safety and Features - OpenAI introduces "safe completions," a new safety training method for GPT-5 [2] - OpenAI will launch four new research preview personalities for all ChatGPT users: Cynic, Robot, Listener, and Nerd [2]