Reproduce “OpenAI GPT-5 System Card” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low and 477 more
GPT-5 is a routed system of fast and reasoning models, with a real-time router selecting between them using signals such as complexity and user intent. Its safe-completions approach aims to keep answers within policy while offering useful help. On HealthBench Hard, gpt-5-thinking scores 46.2%, versus 31.6% for OpenAI o3; these are benchmark results, not evidence of clinical benefit.
Key insights
- A unified model architecture with specialized routers can optimize performance by directing queries to the most suitable component for speed or deep reasoning.
- Shifting from binary safety refusals to "safe-completions" improves both model helpfulness and adherence to safety policies, especially for dual-use prompts.
- Proactive risk management through a "Preparedness Framework" and continuous, multi-faceted evaluations, including external red teaming, is crucial for developing and deploying frontier AI systems responsibly.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.