OpenAI DevDay 2025: Opening Keynote with Sam Altman
Sam Altman's DevDay 2025 keynote introduces major platform updates, including the Apps SDK for building inside ChatGPT, AgentKit for streamlined agent creation, the general availability of the powerful Codex coding agent, and new API access to GPT-5 Pro and Sora 2.
Sam Altman’s keynote at DevDay 2025 began by highlighting OpenAI’s staggering growth since 2023. The developer community has expanded from 2 million to 4 million, weekly ChatGPT users have grown from 100 million to over 800 million, and the API now processes over 6 billion tokens per minute, a significant increase from 300 million. This growth signifies a shift from AI as a novelty to an essential tool for builders. The keynote focused on making it easier for developers to build with AI, centered around four main announcements.
Apps in ChatGPT: The Apps SDK
OpenAI is opening up ChatGPT for developers to build native applications directly within the chat interface. This is enabled by the new Apps SDK, now available in preview. This SDK provides a full stack for developers to connect their data, trigger actions, and render fully interactive UIs.
Key features of the Apps SDK include:
- Built on MCP: The SDK is built on the open standard MCP (Message-Passing and Communication Protocol), ensuring developers have full control over their backend logic and frontend UI.
- Distribution: Apps built with the SDK can reach hundreds of millions of ChatGPT users. Users can invoke an app by name (e.g., “Figma, turn this sketch into a workable diagram”) or discover them through contextual suggestions within a conversation.
- Monetization: Future plans include support for multiple monetization methods, including the new Agentic Commerce Protocol for instant checkout.
- Developer Guidelines: A draft of developer guidelines has been published. Apps meeting these standards will be eligible for listing in a new directory. Apps that meet higher standards for design and functionality will be featured more prominently.
A live demo showcased apps from partners like Coursera, Canva, and Zillow. The demo highlighted the ability to render interactive web content, including videos and maps, directly in the chat. A key feature, “Talking to Apps,” allows ChatGPT to understand the context of what a user is interacting with inside an app’s UI, enabling more intelligent and seamless follow-up questions and actions.
AgentKit: From Prototype to Production
To address the challenges of building and deploying AI agents, OpenAI introduced AgentKit, a complete set of building blocks designed to take agents from prototype to production with less friction.
AgentKit consists of three main components:
- Agent Builder: A visual interface for designing the logic, steps, and flows of an agent. It’s built on the familiar Assistants API and allows developers to model complex workflows by connecting nodes for tools, human-in-the-loop interventions, guardrails, and logic.
- ChatKit: A simple, embeddable, and customizable chat interface that developers can integrate into their own applications. This allows for bringing branded chat experiences, powered by AgentKit workflows, directly to users.
- Evals for Agents: New features for measuring agent performance, including trace grading to understand step-by-step decisions, datasets for assessing individual nodes, automated prompt optimization, and the ability to run evaluations on external models.
A live demo showed an agent being built and deployed for the DevDay website in under 8 minutes. The process involved visually wiring up specialized agents, adding tools like file search, creating custom UI widgets, implementing pre-built guardrails for PII masking, and deploying the agent to the live site using a single workflow ID and the ChatKit React component.
Codex: The Future of Writing Software
Codex, OpenAI’s software engineering agent, is now Generally Available (GA). It has evolved into a capable collaborator that works across a developer’s IDE, terminal, GitHub, and in the cloud.
Key updates to Codex include:
- GPT-5-Codex Model: Codex is now powered by a new, purpose-trained version of GPT-5. This model excels at complex tasks like code refactoring and review and can dynamically adjust its thinking time based on task complexity.
- Rapid Adoption: Since August, daily messages with Codex have increased 10x, and the
GPT-5-Codexmodel has served over 40 trillion tokens. Internally at OpenAI, engineers using Codex complete 70% more pull requests. - New Features for Teams:
- Slack Integration: Ask Codex to write code or answer questions directly from Slack.
- Codex SDK: Extend and automate Codex in a team’s custom workflows.
- Admin Tools: Enterprise-grade controls for environment management, monitoring, and analytics.
The demo for Codex showcased its agentic capabilities by building a software application from a simple sketch to control a live camera and the venue’s lighting system. The task was accomplished using voice commands, a wireless game controller, and real-time API integration, all without the presenter writing any code by hand. A final, impressive step showed the voice agent using the Codex SDK to reprogram the application on the fly, adding a “movie credits” overlay in response to a spoken request.
Model and API Updates
Several major model and API updates were announced to support the new tools:
- GPT-5 Pro in API: OpenAI’s most intelligent and accurate model is now available to all developers via the API, ideal for complex domains like finance, legal, and healthcare.
- GPT-Realtime-Mini: A new, smaller voice model is available in the API. It offers the same quality and expressiveness as the advanced voice model but is 70% cheaper.
- Sora 2 in API (Preview): Developers now have API access to Sora 2, the model powering stunning video generation. This version offers significantly improved controllability, allowing for detailed instructions, statefulness, and synchronized, realistic soundscapes. It can also incorporate real-world assets, like a still image of a dog, into a generated video.