The Long Game

OpenAI just launched Codex Sites. Before reading on, it is worth pausing on what Codex actually is—because most people, including many who use it, are still misreading it.

Codex started as a code-generation model fine-tuned on publicly available code from GitHub repositories—the foundation on which GitHub Copilot was built. It then became an open-source orchestration framework. Then a native desktop app. And now, with Codex Sites, it lets anyone generate a deployed web application from a text prompt—a dashboard, a project tracker, a content planner, a review workspace—hosted on Cloudflare Workers, shareable by URL, with authentication baked in. No SaaS subscription. No development sprint. No IT ticket.

Claude Code followed a parallel trajectory: born as a CLI coding tool, it became the fastest-growing contributor to Anthropic's ARR—while absorbing use cases that have nothing to do with writing code.

Both are systematically misread by those who take their names at face value. The name is an artifact of where these products started. It says nothing about where they are going.

The most revealing number: non-developers now represent roughly 20% of Codex's five million weekly users, growing three times faster than developers. Financial analysts, marketers, researchers, operators.

These are not AI coding tools. They are the emerging runtime for knowledge work.

Computing has always started this way: a tool built for engineers, misread by almost everyone else, until it wasn't. The computer was conceived for calculations—and became the universal medium of work. Too many people are making the same mistake today with Codex and Claude Code.

The Anti-Stripe lens

To make sense of why this matters strategically, I want to recommend a recent piece by Riyaz Shamsudeen in The Ken's Zero Shot column: "Anthropic and OpenAI as Anti-Stripe." It is one of the sharper analyses I have read on where the real platform war is headed—and Codex Sites, published days after, reads almost like an illustration of its thesis.

The argument: while Stripe built its business by increasing the GDP of the internet—empowering applications and transactions—Anthropic and OpenAI increasingly appear to be doing the opposite: absorbing that economic value back into a single runtime. The piece describes this through a precise three-step model—Ingest, Codify, Sever—that maps how a frontier AI platform progressively captures, packages, and ultimately displaces the workflows it initially connected to.

Codex Sites is the "Sever" step made literal.

The workflow was already escaping

Codex Sites did not arrive in a vacuum. Look at what is already happening.

Claude Code automates entire codebases from a terminal. MCP servers expose application logic directly to agents. Browser agents navigate Salesforce, SAP, and Workday without a human clicking through. AI workspaces are becoming the primary surface where work gets done—not the applications underneath them.

Work is moving above the software layer. The workflow is no longer anchored to the application.

This isn't a rupture, but the logical next step of a fifteen-year arc: APIs externalized capabilities, microservices externalized internal architecture, headless systems externalized the user experience. Agents are now externalizing the workflow itself—and with Codex Sites, the application too.

Software synthesis

For twenty years, the path from idea to software followed a long chain. Even with AI-assisted coding, the chain remained intact: developer, code, infrastructure, application. Code was the artifact. Developers were the bottleneck.

Codex Sites compresses that chain. The final artifact is no longer code. It's a working application—generated, hosted, authenticated, shareable, from a conversation. The code becomes an implementation detail. The application becomes ephemeral.

Software synthesis: applications generated on demand, used, shared, discarded, regenerated. Software not as infrastructure, but as materialized intent—and increasingly, as an intermediate material that AI itself produces and consumes: a throwaway interface, a thinking scaffold, discarded once the work moves on.

A nuance matters here. Not all software is becoming ephemeral of course. The backbone holds—infrastructure, systems of record, large-scale process layers. What Codex Sites enables belongs to a different zone: exploratory, ephemeral applications, fast to generate, discarded or absorbed into more durable systems once they prove their value. The plankton, not the fishbone—to borrow Yves Caseau's metaphor.

Beyond Power Apps: consuming the web itself

The enterprise angle is captured by the comparison with Power Apps. Microsoft spent years trying to democratize application creation. Power Apps asked: "What application would you like to build?" Codex Sites asks: "What problem would you like to solve?" One abstraction layer deeper—software becomes a consequence of thinking about work. Power Apps attempted to democratize software creation. Codex Sites may make software creation disappear altogether.

But Codex Sites goes further than the enterprise. It targets small businesses, independent publishers, bloggers, operators who today rely on WordPress, Squarespace, Wix, Webflow—the tools that collectively form the web creation layer Stripe built its business enabling. This is where the Anti-Stripe thesis lands even more: OpenAI is not just building above the enterprise stack. It is trying to consume the economic layer that made the internet valuable in the first place.

Apple used to Sherlock applications—a term that entered the tech vocabulary in 2002, when Apple's own Sherlock search feature absorbed the third-party app Watson and killed its market overnight. Frontier AI labs are now doing the same to entire workflows. And with Codex Sites, to the applications themselves.

A multiplayer board

This is not a two-player game. It is a multiplayer chessboard with overlapping ambitions and converging moves.

Microsoft just unveiled MAI-Thinking-1—a move they could no longer defer, as dependence on OpenAI at the foundation layer had become strategically untenable. Meanwhile OpenAI launches Codex Sites—and Anthropic is doing the same from its side: Claude for Legal, Claude for Healthcare, Claude for Financial Services, Claude for Small Business, Claude Cowork. Both are systematically absorbing the application layer—the one Microsoft spent decades building: Office, Excel, Power Apps, SharePoint, Teams, Dynamics. One player moves toward model independence. The other moves toward workflow ownership. The partnership is dissolving at exactly the moment OpenAI starts competing for Microsoft's core territory. The divorce is final. The consumption phase is beginning. And Codex Sites running on Cloudflare Workers—not Azure—may be one of its most concrete manifestations: OpenAI is deliberately building its application layer on infrastructure that is portable, not Azure-native.

At the same time: Anthropic expands Claude beyond coding into general-purpose knowledge work. Google rebuilds Workspace around Gemini. Salesforce repositions CRM as an agentic, headless environment—after decades of hastily rebranded acquisitions, each sold as a new layer of integration, each ultimately serving the same purpose: expand the distribution surface and stay ahead of commoditization. Notion, Airtable, Retool, Linear—all moving pieces toward the same square.

Not just the model layer, or the infrastructure layer. But the place where work happens. The "operating system" of knowledge work.

What makes this board unusual is that many of the pieces are simultaneously allied and competing.

Google has invested billions in Anthropic—while Anthropic competes directly for the same enterprise knowledge work Google is rebuilding Workspace around. Salesforce runs Agentforce on OpenAI models—while positioning CRM as a rival to the very orchestration layer OpenAI is building. Apple integrates Claude Code and Anthropic into Apple Intelligence—while developing its own on-device runtime that ultimately routes around both. Microsoft has invested $13B in OpenAI—while OpenAI deploys its application layer on Cloudflare, not Azure.

Cloudflare deserves a paragraph of its own. It hosts Codex Sites for OpenAI within its Agent Cloud—a full-stack infrastructure partnership (Workers, Sandboxes, managed Memory, Mesh) that makes it far more than a CDN. It has also acquired Astro, the web framework used by Porsche, IKEA, Webflow and Wix—and one our own Michelin digital teams have built extensively on—giving it control over both the framework and the runtime on which AI-generated sites will run. And it serves Anthropic, Meta, and every other lab on the same infrastructure, betting on the edge layer regardless of who wins the model war. NVIDIA co-launches RTX Spark with Microsoft as the hardware platform for agentic Windows—while simultaneously powering the very edge infrastructure on which OpenAI and others build their agent runtimes.

Covering their bets while attacking the others'—alliances in name, competition in practice.

This is an acceleration of the runtime battle I described earlier this year—except the contest has now moved one layer up, from infrastructure to the application itself.

Different companies. Different moves. Same game.

The terminal is back

There is an older computing model underneath all of this.

Before SaaS, the terminal was the universal orchestration layer. Users didn't live inside applications—they combined tools, moved between systems, orchestrated from a common environment. Then SaaS absorbed the workflow. Applications became the center of gravity.

That gravity is shifting back. The terminal is returning—this time language-driven, reasoning across tools, maintaining memory, acting on behalf of the user. And now also generating the tools themselves.

We are back where we started—except that the terminal now speaks, reasons, and builds.

What enterprises need to get right

For enterprises, two things matter.

First, composability. The Ingest → Codify → Sever model works best on organizations that have already surrendered their architectural sovereignty—those that allowed a single provider to codify their workflows. Those maintaining open, substitutable stacks—where agents run on infrastructure they control, where no single provider owns the environment where work happens—are much harder to Sherlock.

Second, governance—and this is where the deepest paradox lives.

Their SaaS vendors are now being forced to integrate AI on top of existing systems of record—adding a new layer without removing what came before, generating inefficiency at scale. Metered pricing replaces flat subscriptions to protect monetization surfaces that commoditization is eroding. And their captive reflexes prevent them from assembling their own harness around the new cognitive infrastructure—the LLMs, the agent frameworks, the MCP connectors that would let them compose rather than consume—or, in the end, be consumed.

The organizations best positioned to navigate this are those that spent years doing the opposite: decoupling their stacks, building microservice meshes, designing composable components, automating CI/CD—and above all cultivating team cultures capable of absorbing and implementing in short loops whatever emerges. For them, MCP, skills ecosystems, and harness engineering are not new disciplines. They are the same discipline, applied to a new layer.

Every move on this board has one purpose. The organizations that recognize it before the endgame still have choices.

Different moves. Same game.