OpenAI DevDay 2025: ChatGPT as AI OS with in chat apps and commerce

OpenAI’s DevDay announcements this October didn’t just add new features to ChatGPT — they repositioned the product as a runtime, distribution channel, and commerce surface that together look very much like an “AI operating system” for third‑party apps, agents, and transactions.

Background / Overview​

OpenAI unveiled a package of tightly coupled capabilities at DevDay 2025 that, when combined, change how developers build and how users interact with services inside a single conversational surface. The core elements are:

• Apps SDK: a preview SDK that lets third‑party services embed interactive, stateful “mini apps” directly inside ChatGPT conversations.
• App directory & discovery: an in‑chat marketplace where users can discover, enable, and manage apps without leaving ChatGPT.
• AgentKit and agent tooling: frameworks for building multi‑step agent workflows that orchestrate tools and services on users’ behalf while preserving conversational state.
• Instant Checkout + Agentic Commerce Protocol (ACP): an in‑chat checkout experience co‑designed with Stripe and exposed via an open protocol so merchants can accept agent‑initiated orders while remaining the merchant of record.

OpenAI’s pitch at DevDay reframes ChatGPT from a single‑task conversational assistant into a persistent, project‑oriented runtime where apps are first‑class conversational participants. Early launch partners demonstrated the idea: Spotify for playlists, Canva for design, Zillow for real‑estate listings, Figma for collaborative design elements, Coursera for learning content, Expedia and Booking.com for travel booking flows — all rendered and interacted with inline in chat.

---

From Chatbot to “Super App”: what changed and why it matters​

The Apps SDK alters three fundamental constraints that defined prior chatbot integrations.

1) From stateless calls to stateful mini‑apps​

Previous plugin or API integrations typically answered a single question, then returned control to the user. The Apps SDK lets services participate in a sustained, multi‑turn interaction: you can ask Canva to produce a poster, request iterations, and export assets — all while the conversation retains context and the app persists as an interactive element in the thread. That shift is what makes the chat behave more like an application runtime than a single‑purpose assistant.

2) Built‑in discovery and distribution​

Building an app for ChatGPT offers immediate distribution advantages. OpenAI can recommend or surface apps at the point of intent inside the chat, drastically reducing the discovery friction that independent developers normally face through app stores, search, or advertising. This is a deliberate platform play: attract developers with lower distribution costs, then capture a portion of transactional value through commerce features and directory placements.

3) Commerce integrated as a first‑class capability​

Instant Checkout and the Agentic Commerce Protocol (ACP) turn recommendations into transactions without sending users to merchant websites. The protocol is open, tokenized for security, and keeps merchants as the merchant of record while allowing agent‑initiated, in‑chat purchases today with U.S. Etsy sellers and a planned roll‑out to Shopify merchants. Stripe, a core payments partner, has published details of the ACP and co‑designed the checkout flow.

---

The “ChatGPT as an OS” analogy — technical anatomy​

Framing ChatGPT as an “operating system” is a strategic metaphor, but it maps to concrete architecture and product components.

• Kernel = GPT foundation models: the reasoning, multimodal understanding and generation engine that powers the entire runtime.
• System calls = Core APIs (Responses API, Realtime API, etc.): programmatic entry points for apps and agents to use model capabilities.
• Shell = ChatGPT UI (web, mobile, and desktop): the user’s interactive surface and session manager where apps are rendered and invoked.
• Application layer = Third‑party apps via Apps SDK: services that register tools, expose structured inputs/outputs, and return rendered content directly in chat.
• Filesystem / data layer = conversation history + connected data sources: the context store that preserves state and project memory across turns and sessions.

OpenAI explicitly ties this architecture to the Model Context Protocol (MCP) — an open specification for connecting LLMs to external tools — and states that MCP is the backbone that keeps the server, model, and UI in sync for Apps SDK experiences. That alignment with MCP standardizes how apps advertise callable tools and return structured results that ChatGPT can render inline.

---

Commerce and the Agentic Commerce Protocol (ACP)​

Instant Checkout is not a gimmick: it’s a carefully designed commerce primitive with three design goals: security, merchant control, and low friction.

• Payments are tokenized and processed through established payment rails so ChatGPT does not hold raw card data.
• Merchants remain the merchant of record and continue to own fulfillment, returns and customer service. The agent initiates a checkout session that the merchant validates and completes.
• The Agentic Commerce Protocol is open‑sourced under an industry‑friendly license so other AI platforms and merchant systems can interoperate with agentic commerce. Stripe publicly confirmed its co‑development role with OpenAI and the initial Etsy/Shopify rollouts.

Practical implications for merchants and payments teams are immediate: integrate a product feed, implement the ACP endpoints, and validate reconciliation and anti‑fraud controls for agent‑initiated orders. Early merchant partners and payment platforms will shape adoption velocity and the trust model for conversational commerce.

---

Verifying the scale: user and developer reach​

At DevDay OpenAI announced audience and ecosystem metrics that underpin the platform narrative. The company discussed hundreds of millions of weekly users (reported as “700+ million” in August and updated to 800 million weekly active users at DevDay), and claimed millions of developers building with its APIs — numbers that explain why developers and merchants would consider ChatGPT a high‑value distribution surface. These user metrics are independently reported across major outlets and company communications.
Cross‑referencing these figures is important because measurement methodologies can differ. OpenAI’s own public posts in August put ChatGPT on track for 700 million weekly active users, and subsequent DevDay remarks referenced a milestone near 800 million weekly users — a rapid growth trajectory that multiple independent publications corroborated in the days following DevDay.

---

Competition: three distinct philosophies for AI integration​

The Apps SDK and ChatGPT’s platform play create a clearer contrast among how the major vendors are attempting to own the future interface layer.

• OpenAI (ChatGPT): Build the assistant as the platform — apps live inside ChatGPT, discovery happens in chat, and commerce can complete in‑chat. The aim is to make the conversational surface the primary hub for a user’s digital tasks.
• Google (Gemini / Chrome / Workspace): Embed AI into existing product surfaces — Search, Chrome, Android, Workspace — preserving the browser and app paradigms while offering AI enhancements. This model keeps the web and established distribution channels central.
• Microsoft (Copilot & Copilot Studio): Layer AI into productivity applications and operating systems (Windows, Office, Azure) to provide enterprise‑centric, deeply integrated assistant experiences within the tools users already rely on. This plays to Microsoft’s distribution strength in the enterprise.

A fourth player, specialist offerings like Perplexity’s Comet, emphasize transparent sourcing and research accuracy rather than broad app orchestration. Each route is viable and will likely coexist, but the battleground is now the conversational surface and who controls discovery, identity, and payments within it.

---

Strengths of OpenAI’s platform approach​

• Unified UX reduces context switching: Users can research, iterate, and transact in a single conversational thread. That increases convenience and reduces friction for multi‑step tasks.
• Powerful distribution for developers: Access to hundreds of millions of weekly users and in‑chat discovery can materially shorten the path to reach and monetize customers.
• Agentic automation possibilities: Agents plus stateful apps enable higher‑order automations that combine browsing, data plumbing, file manipulation, and checkout — less glue code, more orchestrated value.

---

Risks, trade‑offs and unanswered questions​

Embedding apps, commerce, and identity into a single conversational surface concentrates both power and systemic risk. Key concerns include:

• Data privacy and permissions complexity: Apps can request and act on user data; permission granularity, auditability, and cross‑border data flows will demand a sophisticated governance model. Enterprises must vet how conversation history, connected accounts, and product feeds are stored and exported.
• Fraud and payment security: Agent‑driven checkouts open novel attack vectors (manipulated sessions, token replay, social‑engineering via chat). The ACP and tokenization mitigate some risk, but merchant implementations and monitoring will be critical in practice.
• Hallucinations and data integrity: When an AI “speaks” for an app, errors matter. Systems must ensure the merchant’s authoritative API is the source of truth for any transactional or inventory data to avoid deceptive or incorrect purchases.
• Platform lock‑in and regulatory scrutiny: Centralizing discovery, identity and payments creates potential antitrust scrutiny and raises the bar for portability and fair access. The ACP’s openness helps, but regulators will watch concentration of power in a single conversational layer.
• Quality, moderation and developer governance: If discovery surfaces low‑quality or unsafe apps, user trust erodes. Effective review, monitoring, and transparent ranking policies are essential but operationally expensive.

These are not theoretical: the platform design deliberately acknowledges these trade‑offs and includes guardrails (permission prompts, app review processes, tokenization for payments). Still, the real test is how these systems behave at scale — how fraud rates, dispute resolution times, and content moderation incidents evolve as millions of users and merchants interact via agentic flows.

---

Guidance for developers, merchants and IT leaders​

For companies evaluating or building on the ChatGPT app platform, pragmatic, risk‑aware approaches will win in the near term:

• Start with a single, high‑value, low‑risk use case that clearly benefits from conversation + state (catalog browsing, read‑only recommendations, iterative design tweaks).
• Harden APIs for idempotency, reconciliation and scale — agent‑initiated traffic patterns are different from human browser traffic and will stress checkout endpoints.
• Design clear permission flows and minimal data scopes — ask for the least privilege required and make consent explicit at first use.
• Prepare commerce operations for new support channels, disputes and chargebacks; define SLA and escalation paths for agent‑initiated purchases.
• Maintain a multiplatform strategy — don’t bet everything on a single conversational surface; preserve exportable data and portable features to reduce lock‑in risk.

---

What to watch next (operational signals)​

Over the next 6–12 months, these metrics will indicate whether the strategy succeeds or stalls:

• App approval throughput and directory quality signals (is OpenAI curating effectively?).
• Instant Checkout adoption rates and merchant dispute/chargeback statistics.
• Fraud incident counts and third‑party security disclosures related to MCP or agent flows.
• Developer churn and monetization uptake — are ISVs finding reliable revenue through the platform?
• Regulatory inquiries or policy changes around conversational commerce and platform dominance.

---

Critical takeaways and a sober conclusion​

OpenAI’s Apps SDK, instant commerce primitives, and agent tooling represent a coherent strategy to make ChatGPT the central orchestration layer for everyday digital tasks. The approach is built on a realistic technical foundation — MCP for instrumenting tools, an Apps SDK that defines callable tools and UI metadata, and ACP to standardize agentic commerce — and it leverages a very large and rapidly growing user base to attract developers and merchants.
That said, this is an architectural pivot that concentrates discovery, identity, data, and commerce around a single vendor’s conversational surface. The convenience and productivity gains are real and immediate. The systemic risks — privacy, fraud, content responsibility, and market concentration — are equally real and require continual, visible mitigation. OpenAI has published tooling, standards and process primitives, but the real world will judge the platform on operational metrics: safety incidents, merchant disputes, developer economics, and regulatory responses.
Finally, some claims in early coverage — including rumors about a consumer AI hardware device and exact user counts — should be treated as provisional until OpenAI publishes firm timelines and audited metrics. Readers and implementers should therefore plan for both rapid innovation and the governance burdens that accompany a platform that blurs the lines between assistant, marketplace, and operating surface.
OpenAI has laid down a bold blueprint: make the assistant the place where apps run, commerce completes, and agents act. For developers, merchants, and IT leaders the imperative is clear — experiment now, design for security and portability, and prepare operationally for agentic commerce — because whether ChatGPT becomes the “operating system” of the future will depend less on product demos and more on months of real‑world use.

---

Source: The New Stack OpenAI Aims To Make ChatGPT the Operating System of the Future