OpenAI's new GPT-5.4 clobbers humans on pro-level work in tests - by 83%

Editor's take

OpenAI's latest internal benchmarks indicate GPT-5.4 significantly outperforms its predecessor, GPT-5.2, demonstrating an 83% improvement on tasks previously requiring human expertise and reducing errors by 18%. This advancement suggests a substantial leap in the practical utility of large language models for professional applications, potentially impacting fields like legal document review, medical diagnosis assistance, and complex coding tasks where accuracy and nuanced understanding are paramount.

The implications of such performance gains extend beyond mere efficiency. If corroborated by external evaluations, GPT-5.4’s enhanced reliability could pave the way for broader AI integration into critical workflows, accelerating productivity and potentially lowering operational costs for businesses. The significant reduction in false claims is particularly noteworthy, addressing a key concern for widespread adoption of LLMs in sensitive domains.

Future scrutiny will focus on independent validation of these internal metrics, particularly across diverse, real-world professional scenarios. Understanding the specific benchmark tasks and the extent to which GPT-5.4 generalizes its improvements will be crucial. Furthermore, observing how competitors like Google's Gemini or Anthropic's Claude respond to this reported performance differential, and whether they can match or surpass these capabilities, will shape the ongoing LLM development race.