New GPT-5.4 clobbers humans on pro-level work in OpenAI's tests - by 83%
GPT-5.4 matches or outperforms human professionals 83% of the time across nine industries and 44 occupations, with 18% fewer errors and 33% fewer false claims than GPT-5.2.
Exclusive: Future OpenAI models likely to pose "high" cybersecurity risk, it says
AI models are rapidly gaining cybersecurity capabilities, enabling longer autonomous attacks and prompting increased industry preparation for models that could reach high cybersecurity risk.
Meta's Vibes produced unfocused, derivative AI-generated short videos while OpenAI's Sora delivered more compelling, viral outputs due to stronger underlying models.