OpenAI Dev Day 2025: What Just Changed in AI

A few hours ago, Sam Altman took the stage at Fort Mason in San Francisco and announced what might be the biggest shift in AI development we've seen in years. If you weren't watching the livestream, here's what you missed—and why it matters.

The Four Big Announcements

Sam Altman structured today's Dev Day around four major themes:

Apps inside ChatGPT - Build native apps that live inside ChatGPT

AgentKit - Build production-ready AI agents in minutes, not months

Codex goes GA - AI coding agent that writes software for you

New Models - GPT-5 Pro API, Sora 2 API, and Realtime Mini

Let's break down each one and what it actually means.

---

1. Apps Inside ChatGPT: The New App Store Moment

What Happened

OpenAI just opened up ChatGPT for developers to build interactive apps inside conversations. Not just chatbots. Full applications. With UI, data, actions—everything.

They launched a brand new Apps SDK (in preview, available today) that gives you:

Full stack control (backend logic + frontend UI)

Built on MCP (Model Context Protocol) - an open standard

Access to 800 million weekly users

Discovery through conversation (users can find your app by asking for it)

Future monetization through Agentic Commerce Protocol (instant checkout inside ChatGPT)

How It Works

The demo showed live integrations with:

Coursera - "Coursera, teach me about machine learning"

→ The app loads inline with video playback, and you can ask ChatGPT questions about what's happening in the video while watching

Canva - "Canva, turn this into a pitch deck"

→ Creates a full slide deck right inside the conversation, editable on the fly

Zillow - "Show me homes for sale in Pittsburgh"

→ Interactive map with listings, filtering, and direct actions—all inside ChatGPT

The magic? ChatGPT can "talk to your apps" through context APIs. It knows what the user is looking at and can answer questions about it in real-time.

What This Means for the World

This is OpenAI's play for distribution. They're creating a new platform—like the iOS App Store or Chrome Web Store—but powered by natural language.

Think about it:

No more searching for apps

No more learning new interfaces

Just ask and the right tool appears

For developers: This is a massive opportunity. You can reach hundreds of millions of users instantly. If a user already subscribes to your product, they can log in from the conversation. No friction.

For businesses: Your SaaS product can now be discovered through conversation. Someone asks ChatGPT a question, and your app gets suggested if it's relevant.

For users: Software just got 10x easier to use. You talk, apps appear, things get done.

Developer Guidelines are published today, apps can be submitted for review later this year, and a browsable directory is coming.

---

2. AgentKit: From Idea to Production Agent in Minutes

What Happened

Building AI agents is hard. You need orchestration, eval loops, tool connections, UI—and most agents never make it to production.

OpenAI just solved this with AgentKit: a complete set of building blocks to take agents from prototype to production, fast.

AgentKit includes:

1. Agent Builder

Visual workflow designer. Drag and drop nodes (tools, logic, guardrails, human-in-the-loop). Test flows. Ship ideas. No code required to start.

2. ChatKit

Embeddable chat interface for your apps. Fully customizable with your brand. Works across all agent nodes.

3. Evals for Agents

Trace grading (step-by-step agent decisions)

Datasets for testing individual nodes

Automated prompt optimization

Run evals on external models inside OpenAI's platform

4. Connector Registry

Securely connect agents to internal tools and third-party systems through an admin panel.

The Demo That Blew My Mind

Christina from the OpenAI team built and deployed a fully functional AI agent in 8 minutes live on stage. Here's what she did:

Created a "Dev Day Assistant" agent in the visual workflow builder

Added a classifier to route between specialized agents

Built a "Session Agent" with file search tools and a custom widget

Added a "General Dev Day Agent" that speaks like Froge (their mascot)

Added PII guardrails for security

Previewed the workflow, tested it

Published it with one click

Embedded it into the Dev Day website using ChatKit React component

It went live. In 8 minutes.

The agent could answer questions like "What session should I attend to learn about agents?" by searching through files, using custom widgets to display info, and speaking in Froge's style (with ribbits).

What This Means

For developers: You can now build production-grade agents in hours instead of months. No more wrestling with frameworks, orchestration layers, or eval pipelines.

For enterprises: Deploy AI agents across your organization with confidence. Admin controls, security, monitoring—it's all built-in.

For the world: We're about to see an explosion of AI agents solving real problems. The barrier to entry just dropped to near-zero.

---

3. Codex: AI That Writes Your Software

What Happened

Codex, OpenAI's software engineering agent, just went GA (generally available) after months in research preview.

It now runs on GPT-5-Codex, a version of GPT-5 specifically trained for agentic coding. This model is better at:

Code refactoring

Code review

Dynamic thinking (adjusts reasoning time to task complexity)

Codex works everywhere:

Your IDE (VS Code, etc.)

Your terminal (Codex CLI)

GitHub

The cloud

Your ChatGPT account connects it all.

The Demo That Felt Like Magic

Romain from OpenAI took the stage and built a live camera control system for the Dev Day venue using Codex. In real-time. On stage.

Here's what happened:

Asked Codex to create a control panel interface from a sketch

Asked: "Can you lay out a plan to control a Sony FR7 camera?"

Codex figured out the VISCA protocol (30+ years old), implemented it, built a Node server

Wired up an Xbox controller to control the camera

Integrated OpenAI's Realtime API for voice control

Built an MCP server to control the venue lighting system

Controlled lights and camera with voice commands—live

Example conversation with the AI:

> "What do you see on camera?"

> "I see a large audience in front of you with rows of attendees seated and ready."

>

> "Could you shine the lights towards the audience?"

> "I've just lit up the audience with a nice bright light."

>

> "Can you greet everyone here?"

> "Hello, amazing developers at Fort Mason!"

Then, the finale: Romain asked the voice agent to ask Codex (using the Codex SDK) to add a movie-style credits overlay for the attendees—and it did it. In real-time. While the presentation was running.

What This Means

Software development just changed forever.

You can now:

Turn a sketch into working software (minutes)

Build complex integrations without knowing the protocols (Codex figures it out)

Control physical devices through voice + AI

Have AI reprogram your app in real-time based on user feedback

For individual developers: You're now 10x more productive.

For startups: You can build and ship products at unprecedented speed.

For enterprises: (Like Cisco) You can cut project timelines from weeks to days.

An 89-year-old retiree in Japan taught himself to code with ChatGPT and built 11 iPhone apps for elderly users. Medical students at ASU built a virtual patient app to practice difficult conversations. Versailles Palace visitors can now talk to art and sculptures using OpenAI's Realtime API.

The only limit is imagination.

---

4. Model Updates: More Power, More Modalities

OpenAI also released GPT-5 Pro in the API (their most intelligent model for finance, legal, and healthcare applications) and GPT-Realtime-Mini (70% cheaper voice AI with the same quality). Most exciting for creators: Sora 2 video generation is now available in the API—turn sketches into fully rendered video concepts, generate product ads, and add synchronized soundscapes. As Sam Altman predicted: "Voice is going to become one of the primary ways people interact with AI."

---

What This All Means for the World of AI

1. The Acceleration Is Real

Software that used to take months or years now takes minutes. We just watched someone build a production-ready agent in 8 minutes and a camera control system in 20 minutes.

The idea-to-product timeline has collapsed.

2. AI Has Gone From "Ask" to "Do"

We're past the chatbot era. AI now:

Builds apps

Writes code

Controls physical systems

Creates videos

Makes decisions

Takes actions

This is agentic AI—and it's here.

3. Distribution Just Changed

The new platform wars are starting. OpenAI is positioning ChatGPT as the platform where apps live and get discovered. 800 million weekly users is bigger than most app stores.

Developers who build on this platform early will have a massive advantage.

4. Coding Is Now Natural Language

You don't need to learn protocols, frameworks, or complex systems anymore. You describe what you want. AI figures out how to build it.

This democratizes software creation in a way we've never seen. Anyone with an idea can build.

5. Agents Are Going Mainstream

With AgentKit, the barrier to building production agents just dropped to near-zero. Expect to see:

AI agents in every enterprise workflow

Specialized agents for niche industries

Agents that work together (multi-agent systems)

Agents that learn from your business

---

The Bottom Line

As Sam Altman said: "This is the best time in history to be a builder. It has never been faster to go from idea to product."

He's right.

The only question now is: What are you going to build?

---

At Arkhai Technologies, we're already integrating these new capabilities into solutions for Indian businesses. Want to explore how AI agents, Codex, or the Apps SDK can transform your workflow? Let's talk.