OpenAI introduces GPT-5

OpenAI recently introduced GPT-5 as the most advanced AI system so far, with a significant leap in intelligence over previous models. It featured state-of-the-art performance in coding, math, writing, health, visual perception, and more. GPT-5 was described as a unified system that could determine when to respond quickly and when to spend more time reasoning to give expert-level responses. It became available to all users, with Plus subscribers receiving more usage and Pro subscribers gaining access to GPT-5 Pro, which offered extended reasoning for more comprehensive and accurate answers.

OpenAI introduces GPT-5

GPT-5 was designed as a unified system with a main model for most questions, a deeper reasoning model called GPT-5 thinking for harder problems, and a real-time router to decide which to use based on conversation type, complexity, tool needs, and explicit user intent. The router was continuously trained on real-world usage signals, such as when users switched models or rated responses, and improved over time. When usage limits were reached, smaller mini versions handled remaining queries. Future plans included integrating all capabilities into a single model.

GPT-5 outperformed earlier models on benchmarks, reduced hallucinations, improved instruction following, and minimized sycophancy. It excelled in three major areas of ChatGPT usage: writing, coding, and health. In coding, it became the strongest model yet, showing major improvements in complex front-end generation and debugging large repositories. It could design aesthetically pleasing websites, apps, and games from a single prompt, demonstrating better understanding of design elements such as spacing, typography, and white space. Examples included games like Rolling Ball Minigame, Pixel Art, Typing Game, Drum Simulator, and Lofi Visualizer.

In writing, GPT-5 could turn rough ideas into compelling, well-structured content with literary depth and rhythm. It handled structural ambiguity well and improved everyday writing tasks such as drafting and editing reports, emails, and memos. In health, it scored significantly higher than any previous model on HealthBench, providing more precise, proactive, and context-aware responses. It was designed to help users understand results, prepare questions for healthcare providers, and weigh options, while not replacing medical professionals.

GPT-5’s improvements were demonstrated through examples such as poetry, wedding toasts, cancer risk understanding, and global tour planning. Compared to GPT-4o, GPT-5’s writing was richer, more detailed, and better at establishing imagery and emotional arcs. It set new performance records in math, coding, multimodal understanding, and health, with GPT-5 Pro achieving state-of-the-art scores in high-level reasoning tasks.

The model showed gains in instruction following, agentic tool use, and multimodal reasoning, making it better at handling evolving tasks, interpreting images, and answering questions from diagrams or videos. On economically important tasks, GPT-5 matched or surpassed expert performance in many fields, including law, logistics, sales, and engineering.

GPT-5 delivered faster, more efficient thinking, performing better than OpenAI o3 while using 50-80% fewer tokens in many tasks. It was trained on Microsoft Azure AI supercomputers and was significantly less likely to hallucinate. In factuality tests, “GPT-5 thinking” had about six times fewer hallucinations than o3. It was also more honest in acknowledging its limitations, with deception rates in production reduced from 4.8% with o3 to 2.1% with GPT-5.

The model adopted a new safety approach called safe completions, which allowed it to give helpful answers while respecting safety boundaries, instead of relying solely on refusals. This method improved handling of dual-use domains like virology and reduced unnecessary refusals. GPT-5 also reduced sycophancy through targeted training and evaluation, cutting overly agreeable responses from 14.5% to under 6%.

Customization improved with better instruction following and the introduction of four new preset personalities, Cynic, Robot, Listener, and Nerd, available for text chat and later for voice. GPT-5 thinking was treated as a high-capability model in biological and chemical domains, with strict safeguards and thousands of hours of red-teaming to minimize risks.

GPT-5 Pro replaced OpenAI o3-pro and was built for the most challenging tasks, offering longer, more efficient reasoning. Experts preferred GPT-5 Pro over GPT-5 thinking in nearly 68% of tests, citing fewer major errors and better performance in health, science, math, and coding.

GPT-5 became the new default in ChatGPT, replacing several older models. It was available to Plus, Pro, Team, and Free users starting August 7, 2025, with Enterprise and Edu access following a week later. Pro subscribers had unlimited access, Plus users enjoyed higher limits, and free users received reasoning capabilities after a gradual rollout, switching to GPT-5 mini once limits were reached.