OpenAI DevDay 2025: ChatGPT Becomes an App Platform

TL;DR: OpenAI just flipped the switch from “assistant” to application platform. Developers can build fully interactive apps inside ChatGPT (via a new Apps SDK), ship production-grade agents with AgentKit, speed software work with Codex (GA), and tap new models: GPT-5 Pro, GPT-Realtime-Mini, and Sora 2 in API preview. The headline demo? Zillow running inline in ChatGPT—map, filters, tours—without leaving the conversation.

The big launches

1) Apps inside ChatGPT (Apps SDK, preview)

ChatGPT is now a place where apps live and run. With the Apps SDK (built on MCP), developers can:

• Render fully interactive UIs inline or fullscreen in a chat

• Connect data and trigger actions (logins supported; monetization coming)

• Get in-chat discovery: ChatGPT can suggest your app contextually

Live demos: Coursera (video + “talk to the app”), Canva (generate posters → pitch deck), Figma (turn sketch → diagram), Zillow (interactive map, filters, request a tour).

Why it matters: This is a new distribution layer—apps meet users at intent time, inside the conversation.

2) AgentKit: from idea to production agents

OpenAI bundled the hard parts of agents into one toolkit:

• Agent Builder: Visual node-based workflow design on top of the Responses API

• ChatKit: Embeddable chat + widgets to drop agents into your product

• Evals for Agents: Trace grading, datasets, prompt optimization, external-model evals

• Connector Registry: Secure, admin-controlled hookups to internal tools & third-party systems

Takeaway: Fewer glue layers, faster iteration, built-in measurement = higher odds your agent reaches production.

3) Codex is GA (powered by GPT-5-Codex)

OpenAI’s coding agent is now generally available with team features:

• Slack integration for codegen/Q&A in channel

• Codex SDK to extend/automate your team’s dev workflows

• Admin & analytics for environment controls, monitoring, dashboards

Inside OpenAI, Codex reportedly reviews most PRs and shortens cycles; the pitch is fewer weeks, more days.

4) New models

• GPT-5 Pro (API): The “most intelligent” model for hard, high-accuracy reasoning (finance, legal, healthcare).

• GPT-Realtime-Mini: Advanced voice quality at ~70% lower cost—aimed at voice-first/productized experiences.

• Sora 2 (API, preview): More controllable video with synchronized, realistic audio; better state retention, framing, and remix controls (length, AR, resolution).

Why it matters: Voice becomes a default interface; video becomes product-ready output, not just marketing flair.

By the numbers (from the keynote)

• 800M weekly ChatGPT users

• 4M developers

• 6B tokens/minute processed on the API

Distribution plus scale = a strong reason to build in ChatGPT, not just for it.

The Zillow moment: chat → browse → act

In the demo, a user asked for homes in Pittsburgh, got a Zillow map embedded in ChatGPT, filtered to “3-bed with a yard,” and asked, “How close is this to a dog park?” ChatGPT combined in-app context with search and kept everything in one flow, including the “request a tour” action.

Why it’s important: It’s a clean blueprint for conversational funnels: intent → result → refine → transact, no tab-hopping.

Why this changes the builder playbook

• Apps: With discovery and (soon) monetization in the loop, ChatGPT starts to look like an OS for intent-driven experiences.

• Agents: Standardized building blocks and evals move agents from hype to reliable, measured systems.

• Coding: Codex shifts engineering from “type code” to “direct outcomes,” integrated with where teams already work.

• Modalities: Realtime voice and controllable video expand UI and output beyond text—richer products, same surface.

What to build next (actionable checklist)

1. Design your in-chat app: Pick one high-friction task where inline UI + chat reduces steps (e.g., configure, preview, and purchase—end-to-end).

2. Map an agent with AgentKit: Start with a single workflow; add Evals (trace grading + pass/fail) before shipping.

3. Roll out Codex to your team: Turn it on in Slack for Q&A, adopt the SDK to automate scaffolds/tests/reviews, and watch the analytics.

4. Prototype voice & video: Use Realtime-Mini for a voice-first front door; test Sora 2 for concept videos or product explainers with audio.

5. Plan distribution: Prepare for the Apps directory and in-chat suggestions—clean onboarding, clear value, sensible pricing.

The bottom line

DevDay 2025 reframes ChatGPT as a platform where users can find, run, and pay for software in the flow of conversation. With AgentKit, Codex GA, and new models, the stack to ship useful AI is more complete—and the path from idea to product is shorter than ever. CHeck agnostic AI Agents Marketplace