GPT-5 System Card - OpenAI (August 2025)

GPT-5, released August 2025, is not a single model but a unified system with a smart fast model, a deep reasoning model, and a real-time router that picks which one to use. It scored 94.6% on AIME 2025 (math) and 74.9% on SWE-bench Verified (coding), and makes 5x fewer factual errors than its predecessor o3.

GPT-5 introduced a novel architecture: instead of a single monolithic model, it uses a trained router that examines each input and decides in real-time whether to use the fast model (for simple queries) or the reasoning model (for complex problems). The router is continuously trained on user signals - model switching, preference data, and correctness measurements.

GPT-5-thinking is the deep reasoning mode that shows step-by-step work. GPT-5.2 followed in December 2025 with further improvements. GPT-5.4 launched March 2026 with native computer use.

The router architecture is a significant shift. Previous models forced users to choose between fast/cheap and slow/capable. GPT-5 makes this automatic, optimizing for both cost and quality per-query. This MoE-at-the-system-level approach may become the standard pattern for frontier AI products.

The 5x reduction in factual errors over o3 is especially meaningful - it directly addresses the hallucination problem that has been AI's biggest trust barrier.

Organization: OpenAI. GPT-5: August 2025. GPT-5.2: December 2025. GPT-5.4: March 2026. AIME 2025: 94.6%. SWE-bench Verified: 74.9%.

OpenAI: GPT-5 System Card - https://cdn.openai.com/gpt-5-system-card.pdf

arXiv: GPT-5 System Card - https://arxiv.org/abs/2601.03267

OpenAI: GPT-5.2 update - https://cdn.openai.com/pdf/3a4153c8-c748-4b71-8e31-aecbde944f8d/oai_5_2_system-card.pdf