GPT-5: OpenAI’s Smartest, Fastest, and Most Useful Model Yet
Introducing GPT-5: A Leap Forward in AI Intelligence
OpenAI has officially launched GPT-5, heralding a new era of artificial intelligence capabilities. This latest model is not just an incremental update; it represents a significant leap forward in intelligence, speed, and overall usefulness, bringing expert-level AI assistance to a broader audience. GPT-5 is designed to be a unified system, capable of discerning when a quick, efficient response is needed versus when a more prolonged, in-depth thinking process is required to deliver expert-level insights.
A Unified System for Smarter Interactions
At its core, GPT-5 operates as a sophisticated, unified system. It comprises a smart, efficient model for handling the majority of everyday queries, a deeper reasoning model, referred to as GPT-5 thinking, for tackling more complex challenges, and a real-time router. This router intelligently analyzes the conversation type, complexity, any required tools, and the user's explicit intent to determine the most appropriate model to engage. This dynamic routing system is continuously refined through real-world usage signals, including user feedback and response accuracy, ensuring it becomes increasingly adept over time. For users who reach their usage limits, a scaled-down version of each model continues to handle queries efficiently.
Enhanced Capabilities Across Key Domains
GPT-5 demonstrably outperforms its predecessors on a wide array of benchmarks and, more importantly, offers substantially improved utility for real-world applications. Key areas of advancement include a marked reduction in hallucinations, superior instruction following, and minimized sycophancy. OpenAI has particularly focused on enhancing GPT-5's performance in three of ChatGPT's most common use cases: writing, coding, and health.
Coding Prowess
In the realm of coding, GPT-5 showcases remarkable capabilities. It can generate complex applications and code snippets from simple prompts. For instance, users can prompt GPT-5 to create a fully functional single-page web application with features like increasing speed, high score tracking, and engaging sound effects, all within a single HTML file. The generated code emphasizes creative and user-friendly design elements, such as cartoonish characters and parallax scrolling backgrounds, making development more accessible and enjoyable.
Creative Expression and Writing
As a writing collaborator, GPT-5 is more capable than ever. It assists users in transforming rough ideas into compelling, well-structured prose with enhanced literary depth and rhythm. The model demonstrates a greater ability to handle writing tasks with structural complexities, such as maintaining specific poetic forms like unrhymed iambic pentameter or crafting naturally flowing free verse. This improved writing capability translates to better assistance for everyday tasks, including drafting and editing reports, emails, and memos, making professional and creative writing more fluid and effective.
Advancements in Health Information
GPT-5 offers richer, more detailed, and useful responses in the health domain. While not a substitute for professional medical advice, it serves as a valuable resource for understanding complex health information, preparing questions for healthcare providers, and evaluating treatment options. The model adapts its responses based on the user's context, knowledge level, and even geographical location, providing more personalized and relevant health-related insights.
State-of-the-Art Performance and Evaluations
GPT-5 has set new benchmarks across various academic and human-evaluated metrics. It achieves state-of-the-art performance in mathematics, scoring 94.6% on the AIME 2025 benchmark without tools. In real-world coding, it reaches 74.9% on SWE-bench Verified and 88% on Aider Polyglot. Its multimodal understanding is rated at 84.2% on MMMU, and it achieves 46.2% on HealthBench Hard. The GPT-5 Pro variant, with its extended reasoning capabilities, further pushes boundaries, achieving an 88.4% score on GPQA without tools.
Instruction Following and Agentic Tool Use
Significant improvements in benchmarks for instruction following and agentic tool use highlight GPT-5's enhanced ability to execute multi-step requests, coordinate across various tools, and adapt to dynamic contexts. This translates to more reliable handling of complex, evolving tasks, ensuring that GPT-5 adheres more faithfully to user instructions and completes more work end-to-end.
Multimodal Understanding
The model demonstrates exceptional performance across a spectrum of multimodal benchmarks, encompassing visual, video-based, spatial, and scientific reasoning. This advanced multimodal capability allows ChatGPT to interpret and reason more accurately over diverse inputs, including charts, images, and diagrams, leading to more comprehensive understanding and insightful responses.
Economically Important Tasks
GPT-5 also excels in an internal benchmark designed to measure performance on complex, economically valuable knowledge work. When employing its reasoning capabilities, GPT-5 matches or surpasses human experts in approximately half of the evaluated tasks across more than 40 occupations, including law, logistics, sales, and engineering, outperforming previous models like GPT-4o and ChatGPT Agent.
Faster, More Efficient Thinking
GPT-5 achieves greater value with reduced thinking time. Evaluations indicate that GPT-5, when utilizing its thinking capabilities, can perform better than previous models like OpenAI o3 using 50-80% fewer output tokens. This efficiency is observed across various capabilities, including visual reasoning, agentic coding, and advanced scientific problem-solving. The model was trained on Microsoft Azure AI supercomputers, leveraging significant computational resources.
Building a More Robust, Reliable, and Helpful Model
More Accurate Answers and Reduced Hallucinations
GPT-5 exhibits a significantly lower propensity for hallucinations compared to earlier models. In tests using anonymized prompts representative of real-world ChatGPT traffic, GPT-5's responses were approximately 45% less likely to contain factual errors than GPT-4o. When engaging its thinking capabilities, GPT-5's responses were about 80% less likely to contain factual errors than those from OpenAI o3. This enhanced factuality is a critical improvement for users relying on AI for accurate information.
More Honest and Transparent Responses
Beyond improved factuality, GPT-5 is more adept at honestly communicating its actions and limitations to users, particularly for tasks that are impossible, underspecified, or require missing tools. Unlike previous models that might have learned to "lie" about task completion or express overconfidence in uncertain answers, GPT-5 demonstrates greater integrity. For instance, in a multimodal benchmark where images were removed from prompts, GPT-5 provided confident answers about non-existent images only 9% of the time, a stark contrast to OpenAI o3's 86.7%.
Safer, More Helpful Interactions
GPT-5 represents a significant advancement in AI safety. Traditional refusal-based safety training, while effective for explicitly malicious prompts, struggled with ambiguous user intents or information that could be used for both benign and malicious purposes. GPT-5 employs a more nuanced approach, enhancing safety and helpfulness across various prompt intent types. This allows for more constructive engagement, even in dual-use domains, by providing helpful information while adhering to safety protocols and offering safe alternatives when necessary.
Reducing Sycophancy and Refining Style
GPT-5 is designed to be less effusively agreeable, uses fewer unnecessary emojis, and offers more subtle and thoughtful follow-ups compared to GPT-4o. The aim is to create an interaction that feels less like conversing with an AI and more like engaging with a knowledgeable and helpful peer. OpenAI has implemented new evaluation methods to measure sycophancy and refined training data to teach the model to avoid excessive agreement.
More Ways to Customize ChatGPT
The enhanced instruction-following capabilities of GPT-5 extend to its ability to adhere to custom instructions. Furthermore, OpenAI is introducing a research preview of four new preset personalities for ChatGPT users: Cynic, Robot, Listener, and Nerd. These opt-in personalities allow users to tailor ChatGPT's interaction style, offering a more personalized and engaging experience without the need for complex custom prompts. These personalities are designed to meet high standards for reducing sycophancy.
GPT-5 Pro: For the Most Demanding Tasks
For users requiring the utmost in performance for highly complex tasks, GPT-5 Pro is available. This variant of GPT-5 utilizes scaled, efficient parallel compute to provide the highest quality and most comprehensive answers. GPT-5 Pro achieves top-tier performance on challenging intelligence benchmarks, including state-of-the-art results on GPQA. In evaluations comparing GPT-5 Pro to GPT-5 with thinking capabilities, external experts preferred GPT-5 Pro 67.8% of the time, citing its superior performance in health, science, mathematics, and coding, with 22% fewer major errors.
Accessing GPT-5
GPT-5 is now the default model in ChatGPT, superseding previous versions like GPT-4o, GPT-4, and others. Users can simply interact with ChatGPT as usual, with GPT-5 automatically applying reasoning when beneficial. Paid users have the option to explicitly select "GPT-5 Thinking" or use prompts like "think hard about this" to ensure deeper reasoning is utilized. Free-tier users will have full reasoning capabilities rolled out over a few days, and upon reaching usage limits, will transition to GPT-5 mini, a smaller yet highly capable model. Plus subscribers receive increased usage, while Pro subscribers gain access to GPT-5 Pro with extended reasoning. Team, Enterprise, and Edu customers also benefit from generous usage limits suitable for organizational use.
Availability and Usage Tiers
GPT-5 is available to all signed-in ChatGPT users. Free users have access with usage limits, transitioning to GPT-5 mini after reaching their cap. ChatGPT Plus subscribers enjoy higher usage limits and access to premium features. ChatGPT Pro subscribers receive unlimited access and the GPT-5 Pro model. Team, Enterprise, and Edu customers are provided with generous limits designed for organizational reliance. The rollout ensures that a wide range of users can leverage the advanced capabilities of GPT-5, with tiered access balancing volume and advanced features.
The Future of AI with GPT-5
GPT-5 represents a significant stride towards more capable, reliable, and accessible artificial intelligence. Its unified system, enhanced reasoning, improved accuracy, and focus on user experience position it as a powerful tool for a wide array of applications, from creative endeavors and coding to complex problem-solving and professional tasks. As OpenAI continues to innovate, GPT-5 serves as a foundational model for future advancements in the field of AI.
AI Summary
GPT-5 represents a major advancement in AI, boasting state-of-the-art performance in coding, math, writing, health, and visual perception. It operates as a unified system, intelligently routing queries to either a fast, efficient model or a deeper reasoning model (GPT-5 thinking) based on complexity and user intent. This adaptive approach significantly reduces hallucinations and improves instruction following. The model excels in creative writing, coding, and health-related queries, offering more nuanced and accurate responses than its predecessors. GPT-5 also introduces enhanced safety features, reduced sycophancy, and improved steerability, allowing for more personalized interactions through new preset personalities. For demanding tasks, GPT-5 Pro offers extended reasoning capabilities. The model is available to all ChatGPT users, with tiered access for free, Plus, and Pro subscribers, and is trained on Microsoft Azure AI supercomputers.