OpenAI announces the launch of GPT-5, hailed as a major leap towards Artificial General Intelligence (AGI) and a significant upgrade over GPT-4, now behaving like a PhD-level expert. The new model is described as more useful, smart, fast, and intuitive, capable of tasks ranging from generating full computer programs on demand to providing expert-level assistance in diverse fields like healthcare, education, and coding.
Key highlights include GPT-5's superior performance in coding (SWEBench, Ader Polyglot), multimodal reasoning (MMU), and mathematical reasoning (Amy 2025), coupled with substantial improvements in reliability and factuality, drastically reducing hallucinations. The model introduces "safe completions" for sensitive queries and leverages synthetic data for advanced training.
Functionality demonstrations showcase its ability to create interactive learning demos (e.g., Bernoulli effect visualization), generate nuanced and emotionally resonant writing, build complex web applications with voice-over features, and act as a highly capable, autonomous "pair programmer" that understands codebases, identifies bugs, and refactors large projects. Enhancements to ChatGPT's voice capabilities (naturalness, speed control), personalization (chat colors, customizable personalities), and memory (integration with Gmail/Google Calendar for schedule management) are also rolled out.
The presentation emphasizes GPT-5's real-world utility, particularly for businesses and developers, with testimonials from companies like Amgen, BBVA, and Oscar, and an emotional account of its life-changing impact on a cancer patient's healthcare journey. The API makes GPT-5, Mini, and Nano models available immediately, with varying pricing and reasoning effort options, and introduces new features like custom tools and tool call preambles.
Summarizing the GPT-5 Launch Event:
0:23 GPT-5 Introduction: OpenAI launches GPT-5, positioning it as a major step towards AGI and a significant upgrade over GPT-4. It's designed to be useful, smart, fast, and intuitive, likened to conversing with a PhD-level expert.
1:43 Software on Demand: GPT-5 can write entire computer programs from scratch, plan events, assist with healthcare decisions, and facilitate learning, marking a new era of "software on demand."
3:07 Enhanced Reasoning: GPT-5 integrates advanced reasoning capabilities, automatically thinking deeply when needed to provide comprehensive and accurate responses, combining the best of fast and thoughtful outputs.
4:46 Benchmark Excellence: The model sets new performance highs across various academic evaluations, including SWEBench for software engineering, Ader Polyglot for multi-language coding, MMU for multimodal reasoning (outperforming human experts), and Amy 2025 for mathematical reasoning.
6:12 Improved Reliability and Factuality: A key priority for GPT-5 was reducing hallucinations. It is presented as the most reliable and factual model to date, performing exceptionally well on health-related questions.
7:16 Broad Availability: GPT-5 is rolling out to free, Plus, Pro, and Team users immediately, and to Enterprise and Edu users the following week. Even the free tier will have access to GPT-5, transitioning to GPT-5 Mini (which outperforms GPT-3) after a usage limit.
8:59 Interactive Learning Demos: GPT-5 can explain complex concepts like the Bernoulli effect and generate interactive SVG visualizations or write front-end code (React, Tailwind) for engaging educational tools, making learning more approachable.
15:01 Superior Writing Quality: GPT-5 significantly improves writing quality, offering more rhythmic, nuanced, personalized, and emotionally resonant responses that feel less "AI-like" compared to previous models.
17:27 Advanced Coding Capabilities: Demonstrated by creating a full French learning web application with interactive flashcards, quizzes, and a voice-over game from a detailed prompt. GPT-5 generates hundreds of lines of functional code rapidly and can iteratively fix "rough edges."
23:32 Voice Enhancements: ChatGPT's voice model is improved to sound more natural, integrates video input, and offers consistent cross-language translation. Users can customize voice responses (e.g., concise, single word) and practice language learning with adjustable speaking speeds.
27:12 Personalization and Memory: New features include customizable chat colors, experimental "personalities" (supportive, sarcastic), and enhanced memory. GPT-5 can now integrate with Gmail and Google Calendar (for Pro users first) to understand user context and assist with scheduling and daily tasks.
30:06 Safety and Deception Mitigation: GPT-5 is designed to be significantly less deceptive than previous models. It introduces "safe completions" that aim to maximize helpfulness within safety constraints by partially answering sensitive questions or providing safer alternatives, creating a more robust safety system.
33:03 Novel Training Techniques: OpenAI used GPT-3 to create a high-quality "synthetic curriculum" to train GPT-5 on complex topics, focusing on teaching the "right kind of data" rather than just more data, foreshadowing a recursive self-improvement loop for future AI.
34:32 Healthcare Impact: GPT-5 is highlighted as the best model for health, scoring high on Healthbench. A user shares how ChatGPT helped her understand complex cancer diagnoses, make critical treatment decisions, and regain agency in her healthcare journey. GPT-5 functions as a "thought partner" that anticipates needs.
41:11 Developer-Focused API: GPT-5, Mini, and Nano models are available in the API, with flexible "reasoning effort" parameters. New API features include "custom tools" (plain text outputs with regex/grammar constraints), "tool call preambles" (model explains actions), and a "verbosity parameter."
45:45 Unprecedented Coding & Agentic Performance: GPT-5 sets new standards in coding benchmarks across multiple languages and is the leading model for "agentic tool calling" (complex problem-solving with multiple tools). It boasts a 400K context window for long-range reasoning.
49:52 The Ultimate Pair Programmer: Demos within the Cursor IDE illustrate GPT-5's ability to fix complex bugs, refactor code, and generate entire front-end dashboards and 3D games from concise prompts. It demonstrates strong autonomy, collaborative communication, and aesthetic preferences.
1:05:41 Codebase Understanding: Beyond writing code, GPT-5 exhibits impressive code reading and understanding capabilities, identifying non-obvious architecture decisions and trade-offs within complex codebases.
1:11:36 Enterprise Transformation: 5 million businesses are already using OpenAI tech. GPT-5 enables deep reasoning for life sciences (Amgen), financial analysis (BBVA, reducing 3-week tasks to hours), and clinical reasoning (Oscar). It will also be available to 2 million US federal employees.
1:14:46 Pricing and Availability: GPT-5 is priced at $1.25/$25 per million input/output tokens. GPT-5 Mini and Nano are faster and significantly more affordable, with Nano being 25 times cheaper than GPT-5.