[Paper Review] OpenAI GPT-5 System Card

Alka Singh, Adam Fry+482 more|arXiv (Cornell University)|Dec 19, 2025

Artificial Intelligence in Healthcare and Education3 citations

TL;DR

GPT-5 introduces fast and thinking model layers with a real-time router, improving safety (safe-completions), reducing hallucinations, and boosting performance across health, coding, and multilingual tasks, with extensive red-teaming and safeguards.

ABSTRACT

This is the system card published alongside the OpenAI GPT-5 launch, August 2025. GPT-5 is a unified system with a smart and fast model that answers most questions, a deeper reasoning model for harder problems, and a real-time router that quickly decides which model to use based on conversation type, complexity, tool needs, and explicit intent (for example, if you say 'think hard about this' in the prompt). The router is continuously trained on real signals, including when users switch models, preference rates for responses, and measured correctness, improving over time. Once usage limits are reached, a mini version of each model handles remaining queries. This system card focuses primarily on gpt-5-thinking and gpt-5-main, while evaluations for other models are available in the appendix. The GPT-5 system not only outperforms previous models on benchmarks and answers questions more quickly, but -- more importantly -- is more useful for real-world queries. We've made significant advances in reducing hallucinations, improving instruction following, and minimizing sycophancy, and have leveled up GPT-5's performance in three of ChatGPT's most common uses: writing, coding, and health. All of the GPT-5 models additionally feature safe-completions, our latest approach to safety training to prevent disallowed content. Similarly to ChatGPT agent, we have decided to treat gpt-5-thinking as High capability in the Biological and Chemical domain under our Preparedness Framework, activating the associated safeguards. While we do not have definitive evidence that this model could meaningfully help a novice to create severe biological harm -- our defined threshold for High capability -- we have chosen to take a precautionary approach.

Motivation & Objective

• Introduce GPT-5 as a unified system with fast and thinking models and a real-time router.
• Describe data, training, and safety-oriented methodologies (safe-completions, refusal moderation, and safeguards).
• Evaluate safety challenges, hallucinations, deception, jailbreaks, and multilingual performance against prior models.
• Detail red-teaming efforts, external assessments, and the Preparedness Framework for high-risk domains (bio/chem, cybersecurity).
• Outline governance, monitoring, and future directions for reducing sycophancy and improving safety.

Proposed method

• Define model taxonomy: gpt-5-main, gpt-5-main-mini (fast models) and gpt-5-thinking, gpt-5-thinking-mini, gpt-5-thinking-nano (thinking models).
• Train reasoning models with reinforcement learning to emphasize safe completions over binary refusals.
• Use a multi-layer defense stack including system, developer, and user messages via an Instruction Hierarchy.
• Evaluate safety via disallowed content, jailbreak robustness, prompt-injections, and hallucination benchmarking (LongFact, FActScore, HealthBench).
• Conduct extensive red-teaming (5000+ hours, 400+ testers) including violent attack planning and prompt-injection testing; assess with external researchers and vendors.
• Apply HealthBench, MMLU multilingual benchmarking, BBQ fairness evaluation, and safety/performance comparatives against GPT-4o and OpenAI o3.

Figure 1 : Factuality on ChatGPT Production Traffic (Browsing Enabled)

Experimental results

Research questions

• RQ1 How do safe-completions compare to traditional refusals in reducing safety failures and increasing helpfulness?
• RQ2 What are the safety and performance trade-offs across gpt-5-main and gpt-5-thinking in real-world tasks (health, coding, multilingual)?
• RQ3 How effective are jailbreaks, prompt injections, and instruction hierarchy mitigations across GPT-5 models?
• RQ4 What is the impact of safety safeguards on deception, sycophancy, and hallucination in reasoning tasks?
• RQ5 How do external red-teaming assessments compare to internal evaluations in identifying system-level vulnerabilities?

Key findings

• gpt-5-thinking and gpt-5-main show improved safety and helpfulness versus prior models, with reduced hallucinations and sycophancy.
• Standard disallowed-content metrics show high safety across models; production benchmarks reveal nuanced improvements and some regressions in hate/harassment categories.
• Hallucination rates drop substantially: gpt-5-main’s factual errors are 26% lower than GPT-4o, and gpt-5-thinking’s are 65% lower than OpenAI o3.
• Sycophancy decreased markedly in online and offline evaluations, with significant reductions relative to GPT-4o.
• HealthBench results show gpt-5-thinking achieving substantially higher health safety and performance than prior models, including large reductions in hallucinations and urgent-situation errors.
• Deception monitoring via chain-of-thought showed lower deception rates for gpt-5-thinking (≈2.1%) vs o3 (≈4.8%).
• Image-input safety and multilingual capabilities (13-language MMLU) show competitive performance against baselines.

Figure 2 : Average Hallucination Rate (Browsing Enabled)