Generation AI

OpenAI Dev Day 2025: Apps SDK, Agent Kit, Sora 2, ChatGPT as operating system

Episode Summary

In this episode of Generation AI, hosts Ardis Kadiu and Petar Djordjevic take you inside OpenAI's third annual Dev Day in San Francisco, breaking down the major announcements that are reshaping how we interact with AI. With ChatGPT now reaching 800 million weekly active users, OpenAI is positioning itself as the operating system of the future. Ardis and Petar, who attended the event in person, discuss three major announcement categories: Apps (native applications running directly in ChatGPT with deep integration), Agent Kit (a visual agent builder with built-in evaluation systems), and new models including GPT-5 Pro, Sora 2 video generation, and cheaper image options. They explore what these changes mean for developers, product builders, and higher education professionals, while sharing their first-hand observations from being in the room with 1,500 developers and AI industry leaders. This episode is essential listening for anyone trying to understand where AI platforms are headed and how to prepare for a future where ChatGPT becomes the hub for all your digital work.

Episode Notes

Dev Day Experience: San Francisco and the AI Ecosystem (00:00:36)

First-time experience attending OpenAI Dev Day in San Francisco with 1,500 attendees
The unique culture of San Francisco's tech scene and AI billboards everywhere
Meeting AI influencers, builders from major companies like Netflix, Facebook, Microsoft
Comparing Element451's AI work against world-class builders and feeling competitive
The optimism and grind culture among new builders and startup founders

The Three Big Announcement Categories (00:06:32)

OpenAI's strategic shift: positioning ChatGPT as an operating system
Three main categories: Apps, Agents, and new Models
ChatGPT reaching 800 million weekly active users (not monthly - weekly)
Processing billions of tokens daily across the platform

Apps in ChatGPT: The Third Try at an App Ecosystem (00:10:05)

Native applications running directly in ChatGPT with deep integration
Evolution from plugins (first attempt) to custom GPTs (second attempt) to Apps SDK (third attempt)
Launch partners: Canva, Booking.com, Expedia, Figma, Spotify, Khan Academy, Instacart, Uber, TripAdvisor
Apps can share context with ChatGPT and return custom UI components
Demo showing Coursera courses, Canva slide creation, and Zillow apartment search all within ChatGPT
Apps SDK will be available to all developers by end of year

The Distribution Flywheel and Vendor Lock-in (00:14:53)

800 million users creates massive distribution leverage for app makers
The more users work inside ChatGPT, the more context gets centralized
This strengthens personalization but also increases switching costs
ChatGPT becoming your memory and general assistant
Discussion of potential for ads and payment systems within ChatGPT
Users becoming more sticky to ChatGPT than to individual app websites

Agent Kit: Visual Agent Builder with Native Evals (00:18:38)

Visual agent builder for orchestrating multi-agent workflows
Chat Kit for embedding chat interfaces into applications
Native evaluation system built directly into the platform
Live demo: building a full agent for Dev Day conference in 8 minutes on stage
Pre-built guardrails for PII data and harmful content
Connections to file search, web search, and external systems via MCP protocol
Similar to tools like Zapier, Make.com, and n8n but with embeddable chat widgets

How OpenAI Uses AI Internally (00:23:44)

OpenAI shared three internal use cases at a breakout session
Go-to-market agent: researches customers before meetings, preps demos, closes the loop after meetings
Support agent: handles customer inquiries at scale (not outsourced, built in-house)
When ChatGPT image generation launched, they got 10 million new users in a day
Built-in evals allow systems to improve themselves over time using thumbs up/down feedback

Evals and Prompt Optimization: The Game Changer (00:25:23)

Evals explained: non-deterministic outputs require grading systems
Evolution from human graders to LLM graders
OpenAI introducing prompt optimization using the GEPA algorithm (Genetic Pareto)
System uses all your data and feedback to automatically improve prompts
Connection to DSPY library and the movement toward automated prompt engineering
Not locking users into OpenAI models - can use any model and send traces to the system
Comparison with LangSmith and other tracing tools

New Models: GPT-5 Pro, Sora 2, and Image Mini (00:33:20)

GPT-5 Pro now available via API (12x more expensive than standard ChatGPT)
Takes minimum 15 minutes per task due to deep reasoning capabilities
Sora 2 and Sora 2 Pro for video generation now in API
Sora app showing amazing video generation capabilities
Demo with UK animation studio showing year-long process compressed to minutes
GPT Image 1 Mini: 80% cheaper for cost-sensitive, high-frequency tasks
Enables personalized images at scale for hundreds of thousands of users
Two-tier Sora workflow: use smaller model to nail the prompt, then Pro for high fidelity

Real-Time Voice Models and Device Strategy (00:40:38)

GPT Real-Time Mini Voice: 70% cheaper with improved quality
Discussion about voice quality expectations and production use cases
Speculation about OpenAI's strategy to get models small enough for on-device deployment
The importance of voice as a natural interface for future applications
Concerns about whether cheaper models sacrifice too much quality

Community Reactions and the Agent Debate (00:43:26)

Mixed reactions to Agent Kit announcements
Two camps: those excited about workflow builders vs. those disappointed it's "old paradigm"
Debate about what defines an "agent" - no consensus in the industry
Comparison with Claude Code's different approach: treating LLM as autonomous human
Discussion of workflow builders vs. true autonomous agents

What This Means for Startups and Builders (00:47:40)

Advice: still build in code, don't rely entirely on Agent Kit for production
Agent Kit good for proof of concept and quick distribution
Will take at least a year for App Store to catch fire with normal users
Opportunity to be early in the ChatGPT App ecosystem
Importance of building expertise with OpenAI's tooling and platform

The Everything App and Multi-Platform Future (00:50:30)

ChatGPT positioning as the "Everything App" and operating system of the future
Google announces Gemini Enterprise with similar agent builder capabilities
Q4 2025 prediction: proliferation of agent builders across platforms
Element451's approach: building agents that build agents using conversational interface
Evolution from visual workflow canvas to AI-driven job creation
Proactive AI that evaluates context and takes actions without predefined steps

Final Thoughts: The OpenAI Ecosystem (00:54:13)

OpenAI as one of the most advanced AI labs with 4 million developers on platform
ChatGPT as dominant chat assistant with massive ecosystem impact
Key takeaways from being there in person and seeing the builder community
How these announcements will shape the future of work and higher education