TL;DR

  • Apps inside ChatGPT with a new Apps SDK (built on MCP) let you ship interactive, authenticated web apps directly in ChatGPT.
  • AgentKit debuts: a visual Agent Builder, ChatKit UI, Connector Registry, and agent evals to take agents from prototype to prod.
  • Codex is GA and powered by GPT-5-Codex, with org tooling (Slack integration, SDK, admin/analytics).
  • New models in the API: GPT-5 Pro, Sora 2 (preview) with synced audio, and gpt-realtime updates incl. a cheaper Realtime Mini tier.
  • DevDay 2025 ran Oct 6 at Fort Mason, SF; keynote livestreamed.

Table of contents

  1. What is OpenAI DevDay 2025?
  2. The headline announcements
  3. Deep dives: Apps SDK, AgentKit, Codex GA, Models
  4. Why this changes how you ship software
  5. Fast-start playbooks (code & product)
  6. Pricing/availability snapshot
  7. FAQ

1) What is OpenAI DevDay 2025?

OpenAI’s third DevDay returned to Fort Mason, San Francisco on October 6, 2025, with a keynote stream for everyone and on-site sessions for builders. Expectation-setting posts and previews from OpenAI and the press framed it as a major platform update cycle.

2) The headline announcements

Apps inside ChatGPT (Apps SDK)

Chat experiences now become app surfaces: developers can create interactive UIs that render inline or fullscreen in ChatGPT, connect data/actions, and pass live context back to the model (“talking to apps”). The Apps SDK builds on MCP and ships in preview today. A public directory and monetization (incl. instant checkout) are on the roadmap.

AgentKit

A production-grade agent stack:

  • Agent Builder (visual workflow editor, versionable)
  • ChatKit (embeddable, brandable chat UI)
  • Connector Registry (admin-controlled integrations)
  • Agent evals (datasets, trace grading, prompt optimization; can eval third-party models)
    Available starting today.

Codex (GA) on GPT-5-Codex

OpenAI’s coding teammate is generally available and now runs a GPT-5 variant tuned for refactors, reviews, and long-running tasks—plus Slack integration, Codex SDK, and admin analytics for teams.

New models in the API

  • GPT-5 Pro: high-accuracy reasoning model, now in the API for complex domains.
  • Sora 2 (preview): video + synchronized, realistic sound, better controllability; system card and creator docs out now.
  • gpt-realtime updates (and a cheaper Realtime Mini tier) for production voice agents.

3) Deep dives

Apps SDK: Build apps that live inside chats

  • MCP-based: standardizes how ChatGPT connects to tools & data; Apps SDK extends it to define backend logic + frontend UI.
  • UI surfaces: inline, PiP, fullscreen; can render anything the web can.
  • Context bridge: your app can expose state (“what the user is looking at”) back to the model to make the assistant reactive.
  • Distribution: invoke by name or via suggested apps surfaced in-conversation; directory coming later.
  • Monetization: Agentic Commerce (instant checkout in ChatGPT) is planned.
    Start building in preview today.

Why it’s big: You get built-in distribution to hundreds of millions of weekly users and remove a ton of “glue” work (auth, UI shell, context passing). If your product already has subscribers, they can log in directly in-chat.

AgentKit: From idea → agent → production

  • Design visually, route across tools/agents, add guardrails and human-in-the-loop, then ship.
  • ChatKit drops a polished agent UI into your app; focus on workflows, not scaffolding.
  • Connector Registry gives IT controls and observability.
  • Evals for agents: run trace-level grading, compare nodes/datasets, and even eval external models within OpenAI Evals.
    Available now.

Why it’s big: The stack collapses classic agent pain: orchestration, eval loops, tool wiring, and frontend. Your first useful agent can go live in hours, not weeks.

Codex (GA) on GPT-5-Codex

  • Everywhere you code: IDE, terminal, GitHub, cloud—and now Slack for team workflows.
  • Stronger code intelligence: dynamic “thinking time,” deep reviews, big refactors.
  • Org-ready: admin controls, analytics, and an SDK to automate/extend Codex.

Why it’s big: Teams report faster reviews and higher PR throughput; Codex becomes a persistent teammate that understands your repo and rituals.

New models: GPT-5 Pro, Sora 2, Realtime

  • GPT-5 Pro API: for high-stakes reasoning where accuracy and depth matter.
  • Sora 2: controllable video with synced audio, more realistic physics, broader style range; system card and app docs explain usage & safety.
  • gpt-realtime updates: production-ready voice agents (SIP, image input, MCP server support), with a budget-friendly tier.

4) Why this changes how you ship

  • Distribution is the new moat: Apps SDK gives you immediate reach inside ChatGPT, turning discovery and conversion into part of the chat UX.
  • Agents become productized: AgentKit standardizes what used to be duct-taped—so agents are easier to evaluate, secure, and govern.
  • Engineering leverage: Codex GA + GPT-5-Codex make reviews and refactors faster and deeper across the org.
  • New surfaces: Voice and video creation are now practical for mainstream apps (Realtime + Sora 2).

5) Fast-start playbooks

A) Ship your first ChatGPT app (weekend plan)

  1. Define a narrow job (e.g., “design feedback on Figma frames” or “playlist builder”).
  2. Scaffold with the Apps SDK; add auth for existing subscribers.
  3. Implement context exposure so ChatGPT “knows” the current item/selection.
  4. Add an upsell path; prepare for directory listing once reviews open.

B) Production agent inside your product

  1. Model the flow in Agent Builder (routing + tools + guardrails).
  2. Wrap UI with ChatKit, brand it, and log events.
  3. Add eval datasets and trace grading; iterate prompts with auto-opt.
  4. Wire Connector Registry to your CRM, docs, or data warehouse.

C) Level-up your dev org with Codex

  1. Roll out Slack + IDE extensions; set review rules.
  2. Use GPT-5-Codex for refactors/reviews; track analytics.
  3. Extend with the Codex SDK for repo-specific automations.

D) Voice & video features

  • Prototype a voice agent using gpt-realtime; add phone/SIP if needed.
  • Validate a Sora 2 concept pipeline (shot list → prompt → remix) with audio notes.

6) Pricing & availability snapshot

  • Apps SDK: Preview available now. Directory & monetization to follow.
  • AgentKit: Available today.
  • Codex (GA) with GPT-5-Codex + org features rolling out; see release notes for plan differences.
  • GPT-5 Pro API, Sora 2 (preview), gpt-realtime updates: available; check docs for regions/quotas.

7) FAQ

What exactly are “apps in ChatGPT”?
Web apps that render inside ChatGPT with full UI and backend control, built with the Apps SDK (on MCP). They can be invoked by name or suggested contextually during a conversation.

How do I keep agents safe and measurable?
Use AgentKit guardrails and the new evals (datasets, trace grading, auto-prompt optimization). Admins manage permissions via Connector Registry.

Is Codex just for individuals?
No—Codex is GA with enterprise-grade features: Slack integration, SDK, and admin/analytics for org-wide adoption.

What’s new about Sora 2?
Better physical realism, stronger control/steerability, and synchronized audio—plus an app workflow for creators.

Final take

OpenAI DevDay 2025 wasn’t just model drops; it was a platform turn. Apps live inside ChatGPT (distribution), agents get production rails (AgentKit), engineering workstreams get a teammate (Codex GA), and voice/video become default surfaces (Realtime + Sora 2). If you’re building in Q4, you have everything you need to go from idea → shipped product, fast.

Categorized in:

OpenAI

Tagged in:

, , ,