GPT-5 System Card

Source: OpenAI Date: 2025-08-07 URL: https://openai.com/index/gpt-5-system-card

Summary

The comprehensive safety and capability documentation for GPT-5, published on launch day alongside the enterprise case studies and product announcements. The system card covers capability evaluations across the full Preparedness Framework (CBRN, cyberoffense, autonomy, persuasion), red-teaming results from external evaluators (Metr, ARC, external red-teamers), alignment methodology, and deployment restrictions. GPT-5 is the first GPT-series model where the system card documents AGI-adjacent capability evaluations with genuine concern about positive results.

Implications

The capability threshold thread. GPT-5’s system card is qualitatively different from GPT-4’s because the underlying model is stronger — evaluators are testing thresholds where positive results (e.g., “can the model provide meaningful uplift to CBRN weapons development?”) would trigger deployment restrictions rather than just warning labels. The system card documents whether any such thresholds were triggered and what the risk mitigations are.

The evaluation maturation. By August 2025, system cards include standardized sections tested by third-party evaluators using pre-registered protocols, direct comparison to Preparedness Framework thresholds, and specific mitigations for each identified risk. The GPT-5 system card represents the state of the art in voluntary frontier AI safety disclosure — what the industry expects, and what regulators are starting to formalize as requirements.