DevDay Keynote — Summary 🎤🚀

Opening highlights

• Location & milestone context
  • DevDay in San Francisco; OpenAI origin.
  • Growth since 2023: 2M → 4M developers, 100M → 800M weekly ChatGPT users.
  • API token throughput: ~300M/min → >6B tokens/min.
• Purpose: make it easier to build with AI. Four focus areas announced.

1) Apps inside ChatGPT — Apps SDK (Preview) 🧩

• What: Build full apps that run inside ChatGPT (interactive UIs, connect data, trigger actions).
• Key features:
  • Full-stack Apps SDK (preview) built on MCP (open standard).
  • Developers keep backend logic & frontend control; return HTML templates to integrate.
  • Apps discoverable in conversations and by name; suggestions surfaced contextually.
  • Monetization planned (including Agentic Commerce Protocol / in-chat instant checkout).
• Benefits: reach hundreds of millions of ChatGPT users, new distribution/discovery channel.
• Live demos / partners shown: Coursera (interactive course/video + context-sharing), Canva (create assets, convert posters → pitch deck), Zillow (interactive map + full-screen app & filters).
• Developer roadmap: preview now → submit apps for review later this year → directory & featured placements; developer guidelines draft published.

2) Agents — AgentKit (GA preview) 🤖🛠️

• What: Complete building blocks to design, deploy, and optimize agents from prototype → production.
• Main components:
  • Agent Builder: visual workflow/node-based editor to design logic, tools, guardrails, human-in-loop.
  • ChatKit: embeddable chat UI component for apps.
  • Evals for Agents: trace grading, node-level datasets, automated prompt optimization, evals across external models.
  • Connector Registry: securely link agents to internal & third-party systems with admin controls.
• Use cases / demos:
  • Albertsons: agent analyzing store sales context and recommending actions.
  • HubSpot: improved Breeze assistant combining knowledge base + local data + policies.
  • Live demo: Christina built a DevDay site agent in ~8 minutes (workflow, tools, widget, guardrails, deploy → site integration via ChatKit).
• Availability: AgentKit available today; visual builder, publish workflows, run evals; launch partners and examples shown.

3) How we write software — Codex & CodeX Features 🧭💻

• Codex (now GA)
  • Built on GPT-5-Codex (GPT-5 variant fine-tuned for coding & agentic coding).
  • Improves refactoring, code review, reasoning depth; dynamic compute/time for tasks.
  • Usage growth: 10x daily messages since early Aug; >40T tokens served on Codex.
  • Internal impact: engineers using Codex complete ~70% more PRs; many PRs reviewed via Codex.
• New tooling & integrations:
  • Slack integration for Codex.
  • Codex SDK for team workflows & automation.
  • Admin controls, monitoring, analytics for enterprise adoption.
• Live demo (Romain):
  • Built camera control UI, wired VISCA protocol, integrated Xbox controller and real-time voice control — largely generated/implemented by Codex without manual coding.
  • Showed voice agent controlling lights, greeting audience, taking photos.
  • Demonstrated Codex CLI, VS Code integration, ability to create MCP servers, live-reprogram apps via Codex SDK.

4) Model & API updates — GPT5 Pro, Realtime Mini, Sora 2 🎯

• GPT-5 & GPT5 Pro
  • GPT-5 previously released (strong at agent steering & end-to-end coding).
  • GPT5 Pro now in the API — highest reasoning/accuracy model (good for finance, legal, healthcare).
• Voice:
  • gpt-realtime-mini — smaller, cheaper realtime voice model (same quality/expressiveness, ~70% cheaper).
• Creatives / Video: Sora 2 (preview in API)
  • Sora 2 preview available in API for high-quality controllable video + synchronized soundscapes.
  • Controls: length, aspect ratio, resolution, remixing; supports converting stills to animated scenes.
  • Demo: concept-to-ad workflows and toy design (Mattel partnership).

Overarching message & impact ⚡

• Major announcements to accelerate builder productivity:
  • Apps SDK (in-ChatGPT apps & discovery)
  • AgentKit (visual agent builder, ChatKit, evals, connectors)
  • Codex GA + SDK + integrations
  • New API models: GPT5 Pro, gpt-realtime-mini, Sora 2 preview
• Theme: building with AI is faster and accessible — going from idea → product in minutes; less infra/teams needed, focus on ideas.

Actionable next steps for developers 🏃‍♀️💡

• Try Apps SDK preview to build in-ChatGPT apps; follow developer guidelines draft.
• Explore AgentKit to prototype agents visually; use Connector Registry for secure data access.
• Adopt Codex (GA) and its Slack/SDK integrations to speed team workflows.
• Test GPT5 Pro for high-accuracy tasks; experiment with gpt-realtime-mini for voice and Sora 2 preview for video creativity.

Closing

• OpenAI thanked developers and team; encouraged builders to experiment and provide feedback. 🎉