OpenAI GPT-5 System Card

OpenAI has introduced GPT-5, a unified AI system that utilizes a real-time router to dynamically allocate queries between a fast, high-throughput model for general questions and a deeper reasoning model for complex problems. This architecture allows the system to optimize for speed and depth of reasoning based on conversation type, tool requirements, and explicit user intent.

System Architecture and Model Variants

The GPT-5 system is composed of multiple model variants designed for different performance tiers and use cases. A real-time router, trained on user preference rates, correctness, and model-switching signals, decides which model to use. When usage limits are reached, "mini" versions of the models handle the remaining queries.

Model Mapping

Previous Model GPT-5 Model
GPT-4o gpt-5-main
GPT-4o-mini gpt-5-main-mini
OpenAI o3 gpt-5-thinking
OpenAI o4-mini gpt-5-thinking-mini
GPT-4.1-nano gpt-5-thinking-nano
OpenAI o3 Pro gpt-5-thinking-pro

Availability and Access

  • gpt-5-main and gpt-5-main-mini: High-throughput models used for general queries.
  • gpt-5-thinking: The primary reasoning model, available via API and in ChatGPT.
  • gpt-5-thinking-mini: A smaller, faster version of the reasoning model.
  • gpt-5-thinking-nano: A developer-focused, small and fast reasoning model provided via API.
  • gpt-5-thinking-pro: A version of gpt-5-thinking in ChatGPT that utilizes parallel test-time compute.

Performance Improvements and Capabilities

GPT-5 outperforms previous models on benchmarks and provides faster response times. OpenAI reports significant advances in the following areas:

  • Hallucinations and Instruction Following: There is a reduced rate of hallucinations and improved adherence to instructions.
  • Sycophancy: The system has been designed to minimize sycophancy.
  • Core Use Cases: Performance has been specifically "leveled up" in writing, coding, and health.
  • Safety Training: All GPT-5 models incorporate "safe-completions," a new safety training approach to prevent the generation of disallowed content.

Safety and Preparedness

Under OpenAI's Preparedness Framework, gpt-5-thinking has been categorized as having "High capability" in the Biological and Chemical domain. Although OpenAI states they do not have definitive evidence that the model could help a novice create severe biological harm, they have adopted a precautionary approach by activating associated safeguards.

Sources