# Developer State Of The Union Source: https://www.youtube.com/watch?v=r1R3RDPvPeg Recap page: https://rapidrecap.app/video/r1R3RDPvPeg Generated: 2025-10-08T17:39:00.547+00:00 --- ## Quick Overview OpenAI announced significant advancements across its platform during Dev Day Three, focusing on empowering developers through new models like GPT-5 and Sora 2, expanding Codex capabilities with deep IDE/terminal integration and Multimodal Component Protocol (MCP) support, and launching the Apps SDK for building interactive experiences directly within ChatGPT, all aimed at accelerating software engineering and AI application development. **Key Points:** - OpenAI's initial 2015 plan included solving reinforcement learning and unsupervised learning, which ultimately led to successes like the Dota 2 project and the Unsupervised Sentiment Neuron paper, culminating in GPT-3. - GPT-5 is designed for agentic tasks, excelling at instruction following, tool use, refactoring large code bases, and generating front-end UIs, with Pro versions being smarter but more expensive. - Two new Sora models, Sora 2 and Sora 2 Pro, are now available in the API for high-quality video generation, alongside smaller, cheaper voice (70% less expensive) and image generation models (80% price cut). - Codex now integrates across terminal, IDE, and GitHub, featuring GPT-5 Codex optimized for coding; new team features include a Slack integration and a Codex SDK for embedding the agent into custom workflows. - The Responses API replaces chat completions as the flagship for building agents due to built-in tools (web search, MCP), support for multiple tool calls in one request, and preserving reasoning tokens, making it 20% faster on average. - The new Apps SDK, built on MCP, allows developers to create fully interactive applications inside ChatGPT using familiar web technologies like React, enabling features like controlling venue lighting directly from a chat conversation. - OpenAI released GPT-OSS models, achieving 23 million downloads on Hugging Face, demonstrating community adoption for local or on-premise AI needs, exemplified by projects like OpenSOC and Lifeline Mesh. **Context:** The Developer State of the Union address, presented by Greg Brockman and later by Olivia Gunmore and others, reviewed OpenAI's historical journey from its founding plan in 2015 to current developer tooling. The narrative traced progress from early successes in reinforcement learning (Dota 2) and unsupervised learning (sentiment neuron) to the creation of large models like GPT-3 and the subsequent decision to release technology via API to let developers build applications, leading to today's focus on sophisticated agent building capabilities. ## Detailed Analysis Greg Brockman opened by recounting OpenAI's initial three-step plan: solve reinforcement learning, solve unsupervised learning, and gradually learn complicated things, highlighting early successes in Dota 2 and the Unsupervised Sentiment Neuron, which demonstrated semantic understanding from next-step prediction, paving the way for GPT-3. He emphasized that releasing GPT-3 via API, despite initial skepticism, succeeded because developers connected the technology to real-world applications. Moving to current offerings, GPT-5 Pro is now available in the API, optimized for agentic tasks requiring complex instruction following and tool use, with developers advised to keep prompts simple. Olivia Gunmore detailed the release of Sora 2 and Sora 2 Pro for video generation, along with cost-reduced voice and image models. Codex has been significantly enhanced with GPT-5 optimization, adaptive thinking, and new enterprise features, including Slack integration and MCP tool support (like Figma and Chrome Dev Tools) within the CLI and IDE. The Responses API is now the standard for agents because it natively supports tools and multi-tool calls. Furthermore, Agent Kit provides building blocks for custom agents, demonstrated by RAMP using it to reduce software procurement time from weeks to minutes through visual widget building and trace grading. Finally, the Apps SDK, built on MCP, allows developers to embed fully interactive UIs directly into ChatGPT conversations, enabling rich, visual, and stateful experiences controlled entirely by the user's natural language prompts within the chat interface, with submissions opening later this year. ### Historical Milestones - Initial 2015 plan included solving RL and unsupervised learning - Success seen in Dota 2 reinforcement learning project in 2017 - Unsupervised Sentiment Neuron showed semantics emerging from next-step prediction in 2017 ### Model Updates - GPT-5 is designed for agentic tasks, excels at instruction following and tool use - GPT-5 Pro is smarter but more expensive, reserved for precision cases - Sora 2 and Sora 2 Pro released in API for high-quality video generation ### Codex and Developer Tools - Codex now works across terminal, IDE, GitHub, and cloud with GPT-5 optimization - New features include Slack integration, Codex SDK, and enterprise controls - Codex supports Multimodal Component Protocol (MCP) tools like Figma and Chrome Dev Tools ### Agent Infrastructure - Responses API is the flagship for agents, offering built-in tools and faster performance than chat completions - Agent Kit turns primitives into customizable building blocks for complex workflows - RAMP used Agent Kit to automate software procurement, reducing time from weeks to minutes ### Apps in ChatGPT - Apps SDK allows building fully interactive applications inside ChatGPT using MCP backbone - Developers use standard web tech like React; backend can be any supported language - Example showed controlling venue lighting and creating custom music sequences directly from chat prompts