Written by: Techub News Edited

Introduction

At 12 a.m. Beijing time on May 13, OpenAI held a Spring Update live event, officially launching its next-generation flagship large language model GPT-5. The company’s CEO, Sam Altman, personally hosted the opening, positioning GPT-5 as the most significant upgrade since the release of ChatGPT, marking a "major step" on the path to Artificial General Intelligence (AGI). This release not only unveiled astonishing performance data but also showcased GPT-5's "PhD-level expertise" through a series of live demonstrations in coding, writing, learning, and healthcare. Crucially, OpenAI announced that GPT-5 would be available for free users for the first time, along with an API and enterprise-level solutions for developers, indicating an accelerating era of democratized AI capabilities and deep empowerment across various industries.

Summary

• GPT-5 is positioned as a "PhD-level expert," setting new records across benchmarks such as coding (SWEBench), multidisciplinary reasoning (MMMU), and mathematics (AIME), signifying a significant leap in intelligence.
• The core breakthrough of the new model lies in its "deep reasoning" ability, which can automatically assess problem complexity and think through it, generating more comprehensive and accurate answers, especially excelling in handling open-ended and complex tasks.
• GPT-5 begins rolling out today, including for free users, marking the first time OpenAI's most advanced model is accessible to the free tier, while Plus, Team, and Enterprise users benefit from higher limits or unlimited use.
• ChatGPT receives several major updates: a new more natural voice model, integration with Gmail/Google Calendar, enhanced personalized memory, adjustable "personalities," and interface colors, making AI more personalized and practical.
• At the API level, GPT-5 and its lightweight versions (mini, nano) are being launched simultaneously, introducing features such as "custom tools," "structured output constraints," and "thought chain lead instructions," aimed at becoming the best coding and agent model for developers.

"PhD-level Expert" Emerges: The Performance Leap of GPT-5

In his opening speech, Sam Altman used a vivid analogy to describe the evolution of the GPT series: GPT-3 was like conversing with a high school student, occasionally insightful but often frustrating; GPT-4 resembled a university student, possessing real intelligence and practicality; while GPT-5 is akin to conversing with a "PhD-level expert", providing on-demand, in-depth professional assistance in any area needed. He emphasized that this is not just an enhancement in conversational ability but the dawn of the "on-demand software" era—GPT-5 can write complete computer programs from scratch or assist in various complex tasks ranging from party planning to interpreting medical reports.

OpenAI’s Chief Research Scientist Mark Chen further explained that reasoning capability is the core of OpenAI’s AGI plan. GPT-5 aims to eliminate users' dilemma between quick responses (standard GPT) and thoughtful consideration (reasoning models), as it will automatically determine and deploy the "right amount" of thought to provide perfect answers. Research scientist Max Schwarzer showcased benchmark test results: GPT-5 reached new heights on SWEBench, which measures real software engineering tasks; on the multidisciplinary multimodal reasoning benchmark MMMU, it surpassed previous models and most human experts; and it also excelled on the AIME 2025 problems. Additionally, to address the long-standing issue of “hallucinations” in large models, OpenAI invested significantly in enhancing factual accuracy, particularly on open-ended and complex issues, with GPT-5 being rated as the most "reliable and fact-based model" to date.

Superintelligence for All: Release Plans and Core Features

The most exciting news for the community came from product manager Rennie Song: GPT-5 will begin rolling out gradually to all users starting that day, including free users. This marks the first time OpenAI's most advanced model is accessible to the free tier. Once free users reach their usage limits, they will switch to a lightweight model that remains powerful (even superior to GPT-4o in many dimensions). Plus, Team, and Enterprise users will enjoy higher usage limits, with paid users benefiting from unlimited access to GPT-5. Also launched is the "GPT-5 Pro extended thinking" option for those needing deeper reasoning.

Along with the integration of GPT-5, the ChatGPT app itself will also see a series of significant updates. Research engineer Ruochen Wang demonstrated the newly upgraded voice model, which significantly increases naturalness and introduces a "learning mode" to guide users in gradually understanding a specific topic. Product leader Christina Kaplan then introduced the upcoming deep personalization features: ChatGPT will gain access to Gmail and Google Calendar (initially for Pro users), enabling it to proactively assist users in planning schedules and managing tasks. Furthermore, users can customize the chat interface's color and choose different "personalities" for ChatGPT (such as more supportive, more professional, or slightly sarcastic), aligning its communication style more closely with user preferences. The memory function has also been enhanced to help ChatGPT better understand user context and long-term goals.

The Revolution of Safety and Training: More Responsible, Smarter Generation

OpenAI is acutely aware that the more capable a model is, the greater its safety responsibilities. Safety researcher Saachi introduced that for GPT-5, the team has completely reformulated the security training methods. Traditional models tend to make binary judgments of “complete refusal” or “complete compliance” to prompts, which perform poorly when faced with cleverly dual-use (can be used for both legitimate and potentially harmful) questions. GPT-5 introduces a "safe completion" new method: the model no longer merely assesses the intent of prompts but attempts to maximize "helpfulness" within safety constraints. This may mean partially answering questions, providing only high-level responses, or giving detailed explanations and safe alternatives in refusals. This makes the model more flexible and robust when handling sensitive but legitimate inquiries.

Research scientist Sebastien Bubeck revealed a key training breakthrough behind GPT-5: utilizing previous generation models to generate high-quality synthetic course data. This isn't just about increasing data volume but creating "correct data" to teach the model complex topics. This intergenerational interaction among models signifies a recursive improvement loop, where the previous generation continually aids in enhancing and generating training data for the next generation. He believes this marks the beginning of AI systems moving beyond traditional pre-training and fine-tuning processes towards a more autonomously evolving future.

The Power to Change Reality: Coding, Healthcare, and Enterprise Applications

The event spent considerable time demonstrating GPT-5's potential to transform specific industries. In coding, researchers stunned the audience with multiple live programming demonstrations. Yan Dubois, with just one prompt, had GPT-5 build a fully functional French learning web application in minutes, complete with vocabulary cards, quizzes, and a "cheese-eating word learning" variant of the snake game. Research engineer Adi Ganesh and solutions architect Brian Fioca demonstrated how GPT-5 can understand complex codebases, autonomously locate and fix bugs, and even generate beautiful and interactive financial dashboards or 3D castle games following concise prompts. Cursor co-founder and CEO Michael Truell testified as a partner, showcasing how GPT-5 quickly comprehended non-obvious architectural decisions and security risks within their codebase in a real working environment.

Healthcare was identified by Sam Altman as one of GPT-5's top use cases. He invited a couple—Filipe and Carolina Millon—to share how they used ChatGPT (and later GPT-5) to cope with Carolina's cancer diagnosis. Carolina described how, in panic, she sent a screenshot of her biopsy report filled with medical jargon to ChatGPT and received a clear explanation instantly, giving her a preliminary understanding while waiting for the doctor for hours. In subsequent treatments, when faced with mixed opinions among doctors on radiation therapy decisions, she used ChatGPT to deeply analyze the pros and cons, understand the risks, and ultimately made a responsible and well-informed decision. Filipe emphasized that AI's promise in healthcare is not just for groundbreaking discoveries, but also about creating smarter, more capable patients who can advocate for themselves. After testing, they found that GPT-5 could not only translate medical terminology but also comprehend the context and intent behind the questions, providing more complete, personalized guidance.

OpenAI platform head Olivier Godement listed early applications of GPT-5 in the enterprise sector: the biopharmaceutical company Amgen using it for complex scientific literature and clinical data analysis in drug design; multinational bank BBVA using it for financial analysis, reducing what originally took three weeks into just a few hours; and health insurance company Oscar Health finding it to be the best model for clinical reasoning. Moreover, OpenAI announced a partnership with the U.S. government, allowing 2 million federal employees to utilize ChatGPT and GPT-5.

Developer Tool: Empowering with New API and Ecosystem

For the developer community, research team lead Michelle Pokrass detailed the comprehensive upgrades to the GPT-5 API. In addition to the flagship GPT-5, lighter and more cost-effective versions GPT-5 mini and GPT-5 nano were simultaneously launched, allowing developers to choose based on performance, latency, and cost trade-offs. The API introduces several powerful new features: “custom tools” enable models to call tools in free-form text, better suited for complex scenarios; “structured output” expands support for strictly controlling model output formats through regular expressions or context-free grammar; “thought chain lead instructions” allow models to explain their plans before executing tool calls, enhancing interpretability and controllability; and the “detail level” parameter enables developers to adjust the succinctness or richness of model responses.

Michelle Pokrass also shared a set of impressive benchmark data: in the Tower Square intelligent agent benchmark measuring real problem-solving ability, GPT-5 achieved a score of 97% (whereas all models scored below 49% just two months ago); there were also significant improvements in instruction-following benchmarks. GPT-5's context length has also been expanded to 400,000 tokens. Chief technology officer Greg Brockman summarized that GPT-5 is not only a leader in various benchmarks but its design focuses on real-world usability and deployability, aiming to seamlessly integrate into developers' everyday workflows.

At the end of the event, OpenAI's research director Jakub Pachocki paid tribute to the team and emphasized that the release of GPT-5 is not just a product achievement but a deepening understanding of deep learning as a "magical technology." He believes that the novel ideas presented in GPT-5 are just the beginning, with a longer journey ahead. With the full launch of GPT-5, an era of "a pocket full of PhD experts" officially begins, and how it will reshape work, learning, and creativity is something everyone will eagerly anticipate.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。