October 7, 2025 | Andrew Spoeth

Yesterday, OpenAI hosted its third DevDay — right here at Fort Mason in San Francisco, just a few blocks from my home. The event marked a major inflection point for how people will build with AI in the months ahead.

This year’s theme was clear: AI has moved from “play with” to “build with.”

And the announcements made that shift tangible.

1. Apps inside ChatGPT

Developers can now build fully interactive apps inside ChatGPT using the new Apps SDK, available in preview.

That means familiar tools like Canva, Coursera, Zillow, Spotify, and Figma can now live natively within ChatGPT. You can ask for them by name, launch them directly, and interact with their interfaces without leaving the conversation.

Discovery happens in-line, and soon, through a full directory.

For developers, this opens a new distribution channel to hundreds of millions of users.

For businesses, it’s a signal that ChatGPT is becoming a platform, not just a product.

Learn more: Introducing Apps in ChatGPT →

2. AgentKit: From Idea to Working Agent

OpenAI also launched AgentKit, a set of visual and API-based tools for building and deploying AI agents. It includes:

• Agent Builder for no-code workflow design
• ChatKit UI for branded in-app chat
• Trace Grading for step-by-step evaluation
• Guardrails and connectors for safety and data access

In short, AgentKit makes it practical for teams to build production-grade agents with measurable performance.

Learn more: Introducing AgentKit →

3. Codex for Teams

Codex, OpenAI’s AI engineering assistant, is now generally available — with features aimed at team environments:

• Slack integration for code generation and Q&A
• SDK for workflow automation
• Admin dashboards for policy and analytics

Codex runs on the new GPT-5-Codex model, purpose-built for complex, multi-hour coding tasks like refactoring and reviews.

If you’re a developer eager to build AI-powered tools like those unveiled at DevDay — from agents to in-chat apps — the Generative AI for Software Developers* course is a perfect next step. It teaches you how to design, fine-tune, and deploy generative models that can power real-world applications, helping you turn OpenAI’s latest capabilities into production-ready solutions.

4. New Models in the API

OpenAI also expanded its model lineup:

• GPT-5 Pro – the most capable reasoning model yet, designed for finance, legal, and healthcare-grade accuracy
• Sora 2 Preview – controllable video generation with synchronized audio
• gpt-realtime-mini – a lighter, cheaper voice model

The Numbers That Tell the Story

• 4 million developers have built with OpenAI
• 800 million people use ChatGPT every week
• 6 billion tokens are processed per minute through the API

The scale is astonishing and so is the momentum. AI is now an everyday production tool, not a weekend experiment.

What does this mean for leaders across the enterprise?

Product: Treat ChatGPT as a distribution channel, not just a productivity tool. If your product benefits from conversation or interactivity, start planning a ChatGPT app.

Go-to-Market: Expect early “app store” dynamics — daily value and habit formation will determine visibility. Ship something genuinely useful, not just a demo.

Operations: AgentKit and evals make internal agents viable. Start with a workflow that’s well-defined and measure success in latency and accuracy.

Content & Design: Sora 2 reduces video prototyping from weeks to hours. Creative teams should begin integrating video briefs that include sound and motion.

Engineering: Codex will change how teams review, refactor, and onboard. Define clear policies for when to trust versus verify AI-generated code.

My Take

Apps in ChatGPT could become the new homepage for everyday tasks, reshaping how people find and use digital tools. AgentKit’s trace grading introduces the missing feedback loop for agent performance, helping teams understand and refine AI decision-making. And for marketing and product leaders, now is the time to set a clear, opinionated roadmap for agents, chat-native experiences, and AI-driven vide - ideally in place by early 2026.

If you’d like a deeper breakdown or want to explore what these updates mean for your organization, reach out.

Which announcement stood out most to you?