ChatGPT Just Became an Operating System

I spent Monday watching OpenAI's DevDay event… and it felt like the holidays came early for geeks like me.

A couple of late nights later, and I've finally processed everything they announced.

To sum it up: OpenAI is transforming ChatGPT from a chat interface into a full operating system where developers can build and deploy third-party apps that run natively inside the ChatGPT experience.

This is the type of platform shift that creates massive opportunities for early movers.

So today, I want to break down my notes:

• The key DevDay announcements that actually matter

• What this changes for AI agents and builders

• My take on the opportunity ahead

—@dharmesh

The Key DevDay Announcements That Actually Matter

Sam Altman opened DevDay by organizing announcements into four categories: apps inside ChatGPT, building agents, writing code, and API updates. All equally important - but here are the highlights from my notes:

Apps Inside ChatGPT (The Operating System Moment)

First, OpenAI launched the Apps SDK, allowing developers to build full applications that run directly inside ChatGPT.

The demos were impressive -- Coursera videos playing inline, Canva generating pitch decks in real-time, Zillow providing full home search experiences. All through conversation.

What makes this particularly powerful is that apps can render "anything you can render on the web," including video. Everything stays within ChatGPT, and users interact through conversation.

And best of all… Apps SDK is built on MCP (Model Context Protocol), which I wrote about recently. If you've already built an MCP server, you can support the Apps SDK immediately. I thought their approach was clever: Just add a “resource” (already part of MCP but not widely used) to return the HTML for your app. In any case, by piggy-backing on MCP, the ecosystem exists before the platform even launches.

It’s available in preview now with app submission and listing coming later this year. (Note: I’m a wee bit worried about what the app submission and review process is going to look like — but will save that for another time).

AgentKit: Build and Deploy Agents in Minutes

OpenAI also introduced AgentKit -- everything you need to build, deploy, and optimize agents:

• Visual Agent Builder

• Performance evaluation tools

• Connectors registry for data access

• Built-in guardrails and PII protection

The live demo was exciting (I love live demos). Christina Huang built and deployed a functioning agent in under 8 minutes on stage -- including creating the workflow, adding guardrails, testing, and deploying live to the DevDay website. She’s a braver soul than I.

Then came a moment that was particularly special for me: Sam demonstrated a HubSpot integration using AgentKit right on the main stage. Seeing two of the most impactful companies in my life -- HubSpot (for obvious reasons) and OpenAI (for less obvious ones) -- featured together was incredibly cool.

This integration showcased exactly what's possible: agents that can access CRM data, marketing tools, and customer information become exponentially more useful. It's not just AI talking to AI -- it's AI having access to real business data and workflows.

Cool thing about their Agent Builder is that once you build your agent and publish it, you can access it via an API too. This opens up the opportunity for us to bring the OpenAI agents right into the Agent.ai agent network — which is now 1,800+ agents strong.

Codex Now Generally Available

Next, OpenAI announced their coding agent Codex moved from research preview to GA. They shared an impressive stat that usage increased 10x since early August. I’m not actually surprised by that. OpenAI got the memo and has been investing significantly and iterating a lot on Codex CLI. I think they want to win over developer love and mindshare from Claude Code (and I’m here for it).

The live demo showed Codex controlling a camera with an Xbox controller, building a voice assistant for lights, and creating overlays -- all in real-time.

Coding agents are quickly becoming a standard tool in the way developers work.

API Updates: GPT-5 Pro and Sora 2

Last (but definitely not least), they released major models to the API:

• GPT-5 Pro: Their most powerful reasoning model, now available for developers (warning: Although powerful it’s really expensive. It’s built for some very advanced use cases, so unless you really need it, stick to GPT-5).

• Sora 2: State-of-the-art video generation, available in API just one week after public launch

This is huge. Developers can now build apps and agents that generate professional-quality video (and yes, I've been staying up late experimenting with what's possible when you combine video generation with agents).

What This Changes for AI Agents and Builders

The DevDay announcements are extremely exciting for the entire AI agent landscape. But given I have a word limit in these newsletters, I’ll trim my notes to what I believe are the three most important:

1. ChatGPT Becomes the Distribution Channel

With 800 million weekly users (up from 700 million just weeks ago), ChatGPT is by far the largest distribution platform for AI applications. Build an agent or app, and you immediately tap into that audience.

When HubSpot launched, we built distribution from scratch by being early to blogging and SEO. Today's agent builders get 800 million weekly users on day one.

2. Agents Get Richer Capabilities Faster

AgentKit is dropping the barrier to building sophisticated agents from weeks to minutes. More importantly, agents can now leverage the entire ecosystem of apps -- Canva, Zillow, HubSpot, and any other app in the platform.

This compounds fast. As more apps join, every agent becomes more capable without additional development work. There’s *so* much to build right now.

3. The Agent Network Effect Begins

OpenAI agents are accessible via API, which means they can integrate into other agent networks like agent.ai.

I wrote about this possibility in my 2025 agent predictions -- and honestly wasn't sure how quickly it would happen. Seeing it announced at DevDay was validating, and I think it will accelerate agent adoption exponentially.

The HubSpot integration Sam featured also demonstrates this perfectly agents that access real business data become exponentially more useful than isolated AI assistants.

My Take on the Opportunity Ahead

I've been in tech long enough to see plenty of platform announcements that promised to "change everything." Most don't - but this one feels different.

1. The timing is right. AI models are finally good enough that building on top of them creates genuinely useful applications. Two years ago, GPT Builder was limited by model capabilities. Now, with GPT-5 Pro and real-time capabilities, the foundation is solid.

2. They built on standards. Using MCP as the foundation was brilliant. Developers who've already built MCP servers (like HubSpot) can immediately support the Apps SDK. The ecosystem exists before the platform even launches -- that almost never happens.

3. The developer experience is thoughtful. I've been using OpenAI's APIs for years to build apps and agents. They genuinely understand developers.

My relationship with OpenAI goes back to early GPT-2 API access, before ChatGPT launched. I sold them chat.com and became an investor. They're HubSpot's primary AI partner.

So yes, I'm biased. But even accounting for that bias, what they announced represents a genuine platform shift.

The opportunity for developers and businesses: Build on this platform now while the ecosystem is forming. The apps and agents that establish themselves early will have significant advantages as ChatGPT's "operating system" matures.

If you're building AI products, now is the time to experiment with what's possible.

—Dharmesh (@dharmesh)