Written by: Techub News Compilation
On April 10, 2025, Beijing time, OpenAI held its second developer conference, DevDay, in San Francisco. CEO Sam Altman delivered the opening keynote, announcing several significant updates aimed at empowering developers worldwide. From transforming ChatGPT into an application platform, to greatly simplifying the agent development process, to making AI the core collaborator in software engineering, this release marks a formal shift of AI from a "conversational tool" to a "buildable, integrable, and actionable" ecosystem. Here are the key points from the speech.
Evolution of ChatGPT: From Chatbot to Application Platform
Sam Altman first reviewed the impressive growth of the OpenAI ecosystem: the number of developers increased from 2 million in 2023 to 4 million, with over 800 million users utilizing ChatGPT weekly, and the API's token processing skyrocketing from 300 million to 6 billion per minute. This reflects a profound transformation of AI from a "novel toy" to a "everyday productivity tool."
To further unleash creativity, OpenAI announced the launch of the Apps SDK, allowing developers to build and distribute fully functional interactive applications within ChatGPT. These are no longer simple plugins or tool calls, but "native applications" with independent front-end interfaces, back-end logic, and data connections.
The Apps SDK is built on the open MCP (Model Context Protocol) standard, giving developers complete control over the front and back ends. Its greatest appeal lies in the ability for applications to reach hundreds of millions of ChatGPT users once launched, thereby gaining an unprecedented distribution channel. In the future, OpenAI will also support monetization methods such as in-app purchases.
Live demonstrations vividly showcased this vision:
- Coursera: Users can directly search and play course videos within ChatGPT conversations. The video will be pinned in a small window, allowing users to watch and ask questions simultaneously, with ChatGPT providing instant answers based on the currently playing content (context obtained via API).
- Canva: After brainstorming names for a fictional dog-walking business "Walk This Wag" in ChatGPT, users can directly instruct the Canva app to generate posters and promotional Pitch Decks, completing all creative processes within the conversation flow.
- Zillow: When ChatGPT suggests expanding the business to Pittsburgh, users can invoke the Zillow app, embedding an interactive map to view listings in the conversation, and further request ChatGPT to filter for "three-bedroom houses with a backyard." The application collaborates deeply with the model to provide a seamless experience.
OpenAI engineer Alexi emphasized during the demonstration that this capability of "dialoguing with applications" is the core magic. Applications are no longer isolated tools but are deeply integrated with ChatGPT's intelligence, understanding user intent, and proactively making recommendations. The Apps SDK is open for preview starting today, allowing developers to begin building, with future submissions for application review and official catalog openings.
Agent Kit: Bringing Agents from Prototype to Production
Agents are AI systems that can understand objectives, use tools, and autonomously complete tasks. Despite their great potential, deploying them into actual production still faces tremendous challenges: from selecting frameworks, orchestrating processes, connecting tools, to building UIs and evaluation loops, each step is fraught with complexity.
To address this pain point, OpenAI launched the Agent Kit. This is a complete toolkit aimed at helping developers quickly bring agent ideas into production, containing three core components:
- Agent Builder: A visual workflow creation canvas. Developers can intuitively design agent logic and test processes by dragging and dropping nodes (such as classifiers, tool calls, and conditional branches) without needing to write complex code from scratch.
- Chat Kit: An embeddable chat interface component. Developers can easily integrate this into their applications, customizing brand and interaction style, providing users with a natural dialogue experience with the agent.
- Evaluation Tools: New evaluation features specifically designed for agents, including tracking scores (to understand agent decisions step by step), dataset testing, automatic prompt optimization, and even support for evaluating external models on the OpenAI platform.
Additionally, through OpenAI's connector registry, businesses can safely connect agents to internal tools and third-party systems.
The live demonstration was incredibly impactful. Engineer Christina built and deployed an intelligent assistant named "Ask Froge" for the DevDay official website in 8 minutes using the Agent Kit. This assistant can understand user queries about the conference schedule, retrieve information from documents, and respond in a branded frog style (Froge), while also preventing the leakage of personally identifiable information (PII) through preset "guardrails." The entire process was fully visualized with no code required, and after construction, it could be embedded for real-time use on the website via the Chat Kit component.
Sam Altman also shared a business case: the large American supermarket chain Albertsons built an in-store management agent using the Agent Kit. When ice cream sales unexpectedly dropped by 32%, staff only needed to ask the agent, which could analyze seasonal factors, historical trends, and so on, providing direct suggestions for adjusting displays or placing local ads, greatly simplifying what would have been lengthy reports and meetings.
Codex: Software Engineering Enters the "No Code" Collaborative Era
"We are entering a new era where the way software is written is fundamentally changed," Sam Altman defined the transformation brought by Codex. Codex is OpenAI's software engineering agent designed to work alongside developers, accelerating software creation.
Previously in research preview, Codex is now officially released in a general version (GA). It has been deeply integrated into developers' workflows, supporting IDEs, terminals, GitHub, and cloud environments, and synchronizes all work through ChatGPT accounts. Its underlying model has been upgraded to the specially trained GPT-5 Codex for coding and agent tasks, performing better in code refactoring, reviews, and dynamically adjusting "thinking time" based on task complexity.
Data demonstrates its success: since early August, the daily message processing volume for Codex has increased tenfold, and GPT-5 Codex has become one of the fastest-growing models in OpenAI's history, processing over 40 trillion tokens. Internally at OpenAI, nearly all new code is written by Codex users, and engineers using Codex have seen a 70% increase in the number of pull requests (PRs) completed weekly.
To serve engineering teams, Codex has introduced three new features: Slack integration (to directly let Codex write code or answer questions in team chats), Codex SDK (to extend and automate Codex in team workflows), and new administrator tools and reporting dashboards.
The live demonstration pushed Codex's capabilities to new heights. Engineer Raman completed a series of astonishing tasks without writing a single line of code:
- By issuing natural language commands, Codex created a web control panel to control a Sony camera above the stage.
- Codex researched and wrote the Node.js server code for controlling the camera through the Visca protocol.
- Instructed Codex to connect an Xbox wireless controller to the control interface, allowing the camera to be manipulated with the controller.
- By integrating a real-time voice API and Agent SDK, achieved voice control of the camera (e.g., "focus on the audience") and the venue light system (by building an MCP server).
- Finally, with a voice command, Codex modified React application code live, generating a movie credits-style overlay for all attendees on-site.
The entire process showcased Codex as a powerful "AI teammate": understanding complex intents, researching protocols independently, invoking tools, and writing and integrating code. As Raman said, "The only limit now is your imagination."
Model Updates: GPT-5 Pro, Sora 2, and Realtime Mini Open API
All powerful applications and agents rely on the support of underlying models. Sam Altman announced several model updates:
- GPT-5 Pro API fully open: This is OpenAI's smartest model to date, designed for complex tasks requiring high precision and deep reasoning, such as in finance, law, and healthcare. It is now open to all developers.
- Realtime Mini: A small, fast, low-cost real-time voice model, achieving voice quality and expressiveness equivalent to the Advanced Voice model released two months ago. OpenAI believes that voice will become one of the main modes of human-computer interaction.
- Sora 2 API preview: The highly anticipated text-to-video model Sora 2 is now open for API preview. Its greatest advancement is enhanced controllability, allowing developers to provide detailed instructions, enabling the model to maintain state and produce stylized, accurate, and well-composed videos. Even more exciting, Sora 2 can generate rich sound effects and background music synchronized with the visuals, not just voice. The demonstration showcased how it transformed a photo of a puppy into a dynamic video of a group of puppies playing, complete with synchronized sounds. Toy company Mattel has quickly utilized its API to transform designer sketches into shareable dynamic concept videos.
At the end of the speech, Sam Altman summarized that software building once took months or years, but now, thanks to AI, it can be accomplished in minutes with a good idea. OpenAI's goal is to become the best building platform for this new era through a series of tools such as the Apps SDK, Agent Kit, Codex, and new models. This release was not only a showcase of existing capabilities but also a powerful call to the future of "AI for all" and "immediate realization of creativity."
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。
