OpenAI has officially launched GPT-5, its most capable AI system to date. Rolling out to all ChatGPT users, the model features major upgrades in intelligence, safety, reasoning, and usability across writing, coding, health, and visual perception. GPT-5 will be the default model for signed-in users, with “GPT-5 thinking” and “GPT-5 pro” offering enhanced reasoning and higher quality responses for more complex tasks.
GPT-5 is a unified system composed of three parts: a fast and efficient base model for general queries, a deeper reasoning model for harder problems, and a real-time router that decides which version to use. The router intelligently adapts based on prompt complexity, user instructions, and tool requirements. If a user hits their usage limit, a smaller GPT-5 mini model continues to serve requests. OpenAI aims to unify these components into a single model in the future.
Compared to earlier models, GPT-5 significantly reduces hallucinations, follows instructions more accurately, and adapts more honestly to difficult or ambiguous requests. For example, in real-world tests, GPT-5’s responses were 45% less likely to contain factual errors than GPT-4o, and “GPT-5 thinking” cut hallucinations by 80% compared to OpenAI o3. The model is also less sycophantic and more forthright about its limitations, delivering clearer, safer, and more helpful replies.
GPT-5’s performance sets new records across a wide range of benchmarks. It scores 94.6% on AIME 2025 for math, 74.9% on SWE-bench Verified for code, and 46.2% on HealthBench Hard for medical questions. It also excels in multimodal understanding with 84.2% on MMMU and achieves 88.4% on GPQA with GPT-5 pro. These improvements translate to real-world gains across domains like law, engineering, logistics, and scientific problem solving.
OpenAI has also introduced new customization features. GPT-5 supports improved instruction following and now includes a research preview of preset personalities: Cynic, Robot, Listener, and Nerd. These help tailor the tone of ChatGPT responses without needing prompt engineering and are designed to reduce over-agreement and improve conversation quality.
On safety, GPT-5 introduces a new “safe completions” training method, replacing simple refusals with responses that are both helpful and constrained within safety boundaries. This includes nuanced handling of dual-use queries and improved honesty in situations where tasks cannot be completed. In high-risk domains like biology and chemistry, GPT-5 thinking is treated as high capability and comes with additional safety mechanisms such as red teaming and always-on classifiers.
For developers and power users, GPT-5 pro replaces OpenAI o3-pro and delivers the best performance in the GPT-5 family. It uses parallel compute during generation to reason more deeply, and experts preferred it over the base GPT-5 67.8% of the time in real-world tasks, especially in domains like science, health, and advanced coding.
GPT-5 is now rolling out to ChatGPT Free, Plus, Pro, and Team users. Pro users receive unlimited access and can use GPT-5 pro. Free users get access to the core GPT-5 model, with usage limits and fallback to GPT-5 mini once limits are reached. GPT-5 will reach Enterprise and Edu customers within a week. In addition, GPT-5 is now available through the OpenAI Codex CLI and API for developers.
