OpenAI Sites: The Founder's Guide to Codex 2026

The practical founder's guide to OpenAI Sites: what Codex's new hosted site builder really is, who it is for, and how it stacks up against every other way to build on the web in 2026.

On June 2, 2026, OpenAI shipped Sites and revealed that Codex had crossed 5 million weekly active users in the same breath - OpenAI. The headlines wrote themselves: "OpenAI launches a website builder," "the Wix killer is here," "Codex builds your app for you." Most of them were wrong about what actually launched.

Here is the problem. Sites is not a consumer website builder you sign up for on a Saturday afternoon. It is a Codex capability, in preview, locked to ChatGPT Business and Enterprise workspaces, and built first for internal dashboards, planners, and work-products that live behind a workspace login - OpenAI Developers. If you are a founder who wants a real, public, ownable company on the web, the most important thing to understand about Sites is what it is genuinely for, and where it quietly fails you.

This guide breaks down exactly what Sites does, the technical mechanics of how a prompt becomes a production URL, the six role plugins and annotations that shipped alongside it, the platform strategy that explains why OpenAI built this at all, and a scored comparison of every serious way to build on the web right now, from Lovable and Replit to Wix, Webflow, and the AI labs themselves. It is written for non-technical founders, so it starts high level and then goes deep.

Contents

1. What OpenAI Sites Actually Is (Read This First)
2. The Field at a Glance: Build-From-Chat Platforms, Scored
3. How Sites Works: From Prompt to Production URL
4. The Rest of the Launch: Six Role Plugins and Annotations
5. Why OpenAI Built This: The Platform Logic
6. The Competitors, Profiled
7. Pricing: What You Actually Pay
8. Where Sites Wins and Where It Breaks
9. What It Means for Founders Building a Real Company
10. The Future Outlook: Agents That Build Businesses

1. What OpenAI Sites Actually Is (Read This First)

The single most useful sentence about Sites comes from OpenAI's own developer documentation: "Sites lets Codex create, save, deploy, and inspect websites, web apps, and games hosted by OpenAI" - OpenAI Developers. Read that again slowly. The subject is Codex, not you. The verb is deploy, not design. And the host is OpenAI, not your own server. Every confusion about this product dissolves once you internalize that Sites is a way for an AI coding agent to publish its output as a living, shareable web page rather than a file you download.

That distinction matters because it changes the entire mental model. A traditional builder like Wix or Squarespace hands you a canvas, a template gallery, and a drag-and-drop editor, and you build a site. Sites hands you a chat box inside the Codex app, you describe a work-product, and an agent builds and hosts it. OpenAI's marketing language calls Sites "a new kind of canvas for your ideas" where Codex turns analysis and plans into dashboards, planners, review workspaces, project boards, galleries, and lightweight tools - OpenAI. The recurring nouns there are revealing. These are internal work artifacts, not public marketing websites or consumer products.

Sites did not arrive alone. It was the headline feature of a broader launch titled "Codex for every role, tool, and workflow," which OpenAI unveiled during an "Intelligence at Work" livestream and which also included a new Annotations feature and six role-specific plugins - The New Stack. The framing was deliberate: OpenAI is repositioning Codex from a tool for software engineers into an enterprise work platform for knowledge workers who never write code.

The numbers OpenAI used to justify that pivot are worth stating, with the appropriate caveat that they are self-reported and not independently audited. The company says Codex now has more than 5 million weekly active users, up more than 6x since the desktop app launched in February 2026 - TechCrunch. It also claims that non-developers (analysts, marketers, operators, designers, researchers, investors, and bankers) now make up about 20% of Codex users and are growing more than three times as fast as developers - VentureBeat. Sites is the feature designed to capture that new majority.

Crucially, availability tells you who this is for. Sites is in preview and currently available only for ChatGPT Business and Enterprise workspaces, with Business workspaces having it enabled by default and Enterprise admins needing to turn it on through role-based access control - OpenAI Developers. It is explicitly not available to Free, Plus, Go, or Pro users at launch - Kingy AI. If you are a solo founder on a personal ChatGPT plan, you cannot use Sites today at all. That single fact reframes most of the launch coverage, and it is the reason this guide spends as much time on the alternatives as on Sites itself.

It helps to know how Codex got here, because the speed is part of the story. OpenAI was comparatively late to enterprise: it only introduced plugin support for Codex in March 2026, after Anthropic had already launched its enterprise agents program in February - TechCrunch. The June 2 launch is the catch-up sprint, and it shipped paired with a promise to fold Codex's capabilities directly into the ChatGPT app over the following weeks, so users no longer have to decide when to use Codex versus ChatGPT - Neowin. Read in that light, Sites is not a standalone bet on website building. It is one feature in a rapid effort to make Codex the default agent for everyone in a company, coders and non-coders alike, and the publishing surface where their work becomes visible to colleagues.

2. The Field at a Glance: Build-From-Chat Platforms, Scored

Before going deeper on Sites, it helps to see the whole field at once, because Sites makes sense only in contrast to the dozen other ways you can now turn a sentence into software. The table below scores every serious build-from-chat platform through one specific lens: how well it serves a non-technical founder who wants to build, own, and ship a real web presence or product. That lens is deliberate. A tool can be excellent for enterprise analysts and still score poorly here, which is exactly what happens to Sites.

The methodology is built from first principles about what a founder actually needs, not from generic feature checklists. Output scope (25%) measures how much of a real company the tool can produce, from a single landing page to a full operating business with payments and a database. Ease for non-technical users (25%) measures whether a business owner with no engineering background can actually drive it. Ownership and portability (20%) measures whether you keep and can move what you build, or whether you are renting it from a host. Price and value (15%) weighs entry cost against what you get. Reliability and production-readiness (15%) measures whether you can put the output in front of paying customers. Each cell shows the score and the evidence behind it.

The result is not an accident, and it is not a knock on OpenAI's engineering. Sites scores last on this specific lens because it is built for a different job. Read across its row: it cannot produce a public ownable product (scope 6), you cannot take what it builds anywhere else (ownership 3), and it is a preview gated behind an enterprise plan (price and reliability both held back). For a financial analyst inside a 5,000-person company, none of that matters. For a founder trying to launch a business they own, all of it does. The platforms at the top of the table, including Lovable, Replit, and full-company builders like Founden, win precisely because they optimize for that founder, not for an enterprise workspace.

One honest caveat on the table: nearly every revenue and user figure here is self-reported by the company, so treat them as directional rather than audited. We covered the full taxonomy of these tools, layer by layer, in our AI website builders market map, and ranked the app-focused subset in our Top 20 AI app builders guide, both of which go deeper on individual products than a single comparison table can.

One theme jumps out if you read straight down the ownership column. The platforms that hand you real, portable code (Cursor at 10, then v0, Bolt, Replit, and Lovable clustered in the 8 range) sit near the top, while the fully hosted platforms that keep your work on their servers (Squarespace at 4, OpenAI Sites at 3) sit at the bottom - v0. That ordering is not a coincidence. For a founder, ownership is not an abstract principle; it is the difference between a business you can sell, migrate, or keep running if a vendor changes its terms, and one that exists purely at a platform's discretion. Hold that column in mind through the rest of this guide, because it is the dimension where the marketing and the reality diverge most sharply, and it is the single biggest reason a feature as capable as Sites lands where it does.

3. How Sites Works: From Prompt to Production URL

Understanding the mechanics of Sites matters even for non-technical readers, because the mechanics are the product. Unlike a drag-and-drop builder where you see every change you make, Sites runs an agent through a multi-step build-and-deploy pipeline that you mostly do not see. Knowing what happens under the hood is the difference between trusting the output and shipping a broken page to a production URL by accident.

The flow starts when you name the @Sites plugin in a Codex prompt and describe what you want. From there, Codex builds the project, validates the build, and then offers a choice: save a deployable version for review, or deploy an approved version to a live URL - OpenAI Developers. The critical detail, repeated in OpenAI's own docs, is that "every Sites deployment URL is a production deployment" - OpenAI Developers. There is no soft staging environment. Saving a version is your only review step before something goes live, which is why the documentation pushes you to save and inspect before you deploy.

Underneath, the hosting is more sophisticated than a static page server, and that is what lets Sites build real apps rather than brochures. Sites runs on Cloudflare Workers infrastructure, requiring projects to build Cloudflare Worker-compatible output as ES modules - VentureBeat. Apps that need to remember data use D1, a relational database, and file uploads use R2 object storage, with the project's linkage and storage bindings stored in a small .openai/hosting.json file - OpenAI Developers. This is genuinely full-stack. Codex can build a game with persistent scoring, an internal tool with workspace authentication, or an app with file uploads, not just a landing page.

A typical .openai/hosting.json file looks like this, which is the kind of plumbing Codex manages for you so you never touch it directly:

{
  "project_id": "<project-id>",
  "d1": "DB",
  "r2": null
}

Sharing is governed by three access modes, and understanding them is essential because they reveal the product's true center of gravity. The modes are admins_only (the owner and workspace admins), workspace_all (every active user in the workspace), and custom (specific users or groups you choose) - OpenAI Developers. Notice what is missing: there is no documented fully public, unauthenticated viewing option. Every audience mode requires a workspace sign-in. The official docs simply instruct you to "set the audience before you share a deployed URL" - OpenAI Developers. A Site is something your colleagues open after logging in, not something a stranger discovers from a Google search.

That single design choice, authenticated-by-default sharing, is the clearest signal of intent. To make the practical use concrete, the best third-party walkthrough offers ready-to-run prompts, such as "@Sites Build an internal project dashboard for the operations team. It should show active projects, owners, priority, blockers, due dates, and status" - Kingy AI. The recommended prompt structure asks you to specify the audience, the core workflow, the data source, sign-in requirements, persistence needs, and access scope. That is a meaningful amount of specification, and it tells you that Sites rewards clear thinking about an internal tool far more than it rewards a vague "make me a website" request.

Two technical gaps round out the picture, and both matter for anyone weighing Sites against a real builder. First, the documentation does not mention custom domains at all, so a Site lives at an OpenAI-issued URL rather than at your-company.com - OpenAI Developers. Second, there are no documented numeric limits: no published quota on the number of sites, the site size, storage, or request rate, which is normal for a preview but a genuine planning risk for anything you intend to scale - Cloudflare. Secrets and environment variables are managed in the Sites panel and must never be committed into the project, and each deployable version is tied to the source Git commit that produced it, which gives you a basic version history even without a true staging environment.

The range of things Codex can build through Sites is wider than the "dashboard" framing suggests, which is both its strength and the source of the confusion. The supported types include content-led websites and landing pages, apps with persistent data through D1, file uploads through R2, workspace-authenticated internal tools, and even apps with public sign-in via an external identity provider - Kingy AI. The guidance even nudges builders toward restraint, cautioning against requesting durable storage unless the hosted site needs to remember real product data. In practice this means Sites is genuinely capable of small real applications, but its defaults, its hosting, and its sharing model all pull toward the internal-tool use case rather than the public-product one. The capability is there; the product's center of gravity is somewhere else.

4. The Rest of the Launch: Six Role Plugins and Annotations

Sites was the most eye-catching feature, but the six role-specific plugins are arguably the more strategically important part of the June 2 launch, because they reveal who OpenAI is actually chasing. Each plugin bundles apps, skills, instructions, and workflows so that, in OpenAI's words, a single install turns Codex into "a specialist for a specific role" with no coding required - OpenAI on X. Together the six plugins ship with 62 popular apps and 110 skills - The Next Web.

The six plugins map directly onto high-value white-collar jobs, which is the whole point. They cover data analytics, creative production, sales, product design, public equity investing, and investment banking - TechCrunch. The integrations inside each plugin are not generic. The data analytics plugin wires in Snowflake, Databricks, Hex, and Tableau. The sales plugin connects Salesforce, HubSpot, Slack, Outreach, and Clay. The public equity investing plugin pulls from Moody's, FactSet, LSEG, S&P, PitchBook, and Hebbia - Shopifreaks. These are the daily tools of analysts, marketers, sellers, and bankers, and OpenAI is positioning Codex as the agent that orchestrates them.

The other plugins fill out the same white-collar map with equally specific integrations. Creative production connects Figma, Canva, Shutterstock, Picsart, and Fal to turn a brief into ad variations and campaign assets; product design bridges Figma and Canva to turn static wireframes into clickable prototypes; and investment banking helps bankers prepare pitch materials and comparable-company analysis from trusted data sources - Shopifreaks. OpenAI also said data analytics drew the strongest early adoption, citing roughly 110% growth in that use case, and that it is building toward an open ecosystem where partners publish their own plugins into Codex and ChatGPT - PYMNTS. The bundling is itself the strategy: instead of teaching the model each profession, OpenAI lets the model orchestrate the apps where that expertise already lives.

The plugin roadmap makes the ambition explicit. OpenAI confirmed that more role-specific plugins are already in development for Corporate Finance, Private Equity Investing, Marketing Strategy, Strategy Consulting, and Legal - The Tech Portal. Bloomberg reported the vertical-expansion strategy is squarely aimed at finance, banking, consulting, and legal services, intensifying competition with Anthropic, which rolled out finance-focused agents earlier in 2026 - Bloomberg. Reporting from the legal-tech world even suggests an offering branded "Codex for Legal" is being staffed up - Artificial Lawyer.

The third piece, Annotations, is the quiet workhorse that makes everything else usable. OpenAI describes it simply: "you point to the exact part you want to refine and tell Codex what needs to change" - OpenAI. Developers already used annotations to refine code and websites; the June launch extends the same in-place editing to documents, spreadsheets, and slides. You can select a chart on a slide and ask for a clearer label, highlight a claim in an investment thesis and ask where it came from, or mark a navigation bar on a Site and change its font - Creati.ai. It is the difference between regenerating a whole document and surgically fixing one paragraph, and for non-technical users it removes the single most frustrating part of working with AI.

Taken together, these three features describe a coherent product thesis rather than three disconnected launches. Plugins give Codex the context and tools of a specific role, Annotations let a human refine the output without restarting, and Sites lets that output be published and shared as a living artifact. OpenAI's Chief Revenue Officer Denise Dresser framed the underlying bet plainly: "AI is becoming capable of doing increasingly meaningful work inside organizations," and the challenge now is integrating it into the workflows that run businesses - TechCrunch.

There is a subtle point hidden in this structure. OpenAI is not trying to make Codex an expert in finance or sales; it is trying to make Codex an orchestrator of the tools that already hold that expertise, and the 62 app integrations and 110 skills are the opening inventory of that orchestration layer - The Tech Portal. For a non-technical user, the practical effect is that Codex starts to feel less like a chatbot and more like a colleague who already has logins to everything and knows the standard playbook for the job. That is a genuinely different product from a website builder. It is the clearest evidence that the June 2 launch is fundamentally about owning knowledge work, with Sites positioned as the surface where that work gets published and shared.

5. Why OpenAI Built This: The Platform Logic

To understand Sites from first principles, ignore the feature list and ask the structural question: what does OpenAI actually want to own? The surface answer is "a website builder." The real answer is the application layer itself, the place where work gets done and value gets captured. Sites is one move in a multi-year campaign to make ChatGPT the surface where everything happens, so that the things people create, buy, and use never leave OpenAI's gravity well.

The evidence for this reading is a clean sequence of launches, each adding a layer to the same platform. In October 2025, OpenAI shipped the Apps SDK, built on the open Model Context Protocol, letting third-party apps run natively inside ChatGPT - OpenAI Developers. In December 2025 it opened a ChatGPT app directory, an app store with Featured, Lifestyle, and Productivity categories - TechCrunch. Through 2025 and 2026 it pushed agentic commerce via the Stripe-co-developed Agentic Commerce Protocol, letting users buy inside ChatGPT - Stripe. And in October 2025 it launched the Atlas browser, putting ChatGPT between users and the entire web - Wikipedia. Sites simply extends the same logic to the publishing layer: the web pages and apps people create now live on OpenAI's surface too.

The scale behind this strategy is what makes it credible rather than aspirational. At DevDay in October 2025, Sam Altman said ChatGPT had reached 800 million weekly active users - TechCrunch. By February 2026 that figure was 900 million weekly users with 50 million paying subscribers and more than 9 million business users - TechCrunch. The company was running at roughly a $25 billion annualized revenue run rate by early 2026 - Sacra. And in March 2026 it closed a record $122 billion funding round at an $852 billion valuation - CNBC. When a platform with that reach adds a publishing layer, it is not testing a feature; it is annexing territory.

Two more data points sharpen the picture. The Sites launch came just three weeks after OpenAI stood up the "OpenAI Deployment Company," a venture with more than $4 billion behind it aimed at embedding OpenAI's tools inside large organizations - TechCrunch. On the commerce side, OpenAI's Stripe-built Agentic Commerce Protocol already lets users buy from Shopify and Etsy merchants without leaving ChatGPT - Stripe. Deployment captures the enterprise relationship, commerce captures the transaction, and Sites captures the work-product. The counterweight worth holding in mind is concentration risk: Menlo Ventures data cited by enterprise analysts shows OpenAI's share of enterprise LLM API spend slipping to about 27%, down from roughly 50% in 2023, as Anthropic climbed toward 40% - Kai Waehner. Hosting your internal tools on OpenAI's surface deepens exactly the kind of single-vendor dependence that many enterprises are now actively working to reduce, which is a real tension inside the strategy rather than a footnote to it.

Altman has been unusually direct about the end state. In an interview he argued that "most people will want to have one AI service, and that needs to be useful to them across their whole life," describing a bundled subscription and a "sign in with OpenAI" identity layer threaded across other services - Stratechery. Commentators have summarized the ambition as becoming "the Microsoft of AI," an operating system built on ChatGPT - Windows Central. Sites is the publishing primitive in that operating system. The reason a Codex Site has no public, unauthenticated URL and no export path is not an oversight. It is the strategy working as designed: artifacts created on the surface stay on the surface.

This is where the first-principles read gets uncomfortable for partners. OpenAI launched Sites with Vercel, Wix, Replit, Lovable, Figma, Webflow, and Emergent as early partners - Neowin. Those same partners are app-layer companies that OpenAI's platform could eventually absorb. One widely shared reaction to the launch put it bluntly: infrastructure companies "eventually try to win the platform game, then they learn and take out all their partners on the app layer" - Digg. Whether that prediction holds is the central tension of the 2026 AI build-tools market, and it is why ownership scored so heavily in the assessment table above.

6. The Competitors, Profiled

The honest way to evaluate Sites is to put it next to the companies that have spent years building exactly what Sites only gestures at: real, ownable, production-grade web products created from natural language. These are also, in several cases, OpenAI's own launch partners, which makes the competitive dynamics genuinely strange. Each profile below covers what the company does, its 2026 scale, its pricing posture, and where it beats or loses to Sites.

Lovable is the breakout star of the category and the clearest counterexample to Sites. The Stockholm-based company, led by CEO Anton Osika, turns plain-text prompts into full-stack apps with authentication, databases, and custom domains. Its self-reported ARR trajectory is staggering: roughly $100M by July 2025, $200M by November 2025, and $400M by February 2026 with just 146 employees - TechCrunch. In December 2025 it raised a $330M Series B at a $6.6 billion valuation, more than tripling its mid-2025 valuation in five months - TechCrunch. Where Sites hosts internal dashboards behind a login, Lovable ships public products you can export to GitHub and own. It is, simultaneously, an OpenAI Sites partner and the most direct refutation of the idea that OpenAI has built a real builder.

The scale of Lovable's run is worth pausing on, because it reframes what is possible for a founder. Osika claims the company reached $100 million ARR faster than OpenAI, Cursor, or any software company in history, and at $400M ARR with 146 staff it runs at roughly $2.77 million in revenue per employee - TechCrunch. The lesson is not the revenue figure, which is self-reported and best treated with caution, but what it implies about the tooling. A 146-person company now serves millions of builders, which means the same engine available to a global enterprise is available to a solo founder for $25 a month. The build layer has been genuinely democratized; the open question is what a founder does with capability that is no longer scarce.

Replit occupies a similar position with a more developer-flavored history. The company raised a $400M Series D at a $9 billion valuation in March 2026, tripling its valuation in six months, and told investors it hopes to reach $1 billion ARR by the end of 2026 - TechCrunch. Its flagship, Replit Agent 4, runs parallel agents on a single project and builds web apps, mobile apps, and data apps. Replit Core costs $20/month and Pro $95/month - Replit. Like Lovable, Replit gives you real, exportable code and a public deployment, which is why both rank above Sites for a founder building something they intend to keep.

Vercel's v0 approaches the same goal from the developer-tooling side. Vercel closed a Series F at a $9.3 billion valuation in September 2025, and v0 had surpassed 4 million users by early 2026 - Vercel. v0 generates frontends and full-stack apps that deploy to Vercel's infrastructure, priced from a $20/month Premium tier on a credit system - v0. It leans technical, which caps its ease-of-use score for non-coders, but its output is genuinely yours and production-grade. Our Top 20 AI app builders guide breaks down how v0, Bolt, and Lovable differ on code ownership in particular, which is the dimension where they most outclass Sites.

Two younger builders show how fast this market moves. Bolt.new, built by StackBlitz on browser-based WebContainers, went from launch to about $40 million ARR in roughly five months and crossed 7 million users by late 2025, with Pro pricing at $25/month on a token system - Sacra. Emergent, led by Mukund Jha, targets non-technical users building full-stack, mobile-first apps through autonomous agents; it reported $100 million ARR and 6 million users across 190 countries by February 2026, and raised a $70 million Series B led by SoftBank's Vision Fund and Khosla - Sacra. Emergent is also an OpenAI Sites partner, which is doubly notable because its apps reportedly run largely on Anthropic's Claude models. The pattern across both should reassure founders rather than worry them: the build layer is crowded, competitive, and getting cheaper by the month, which is exactly the condition under which ownership and distribution, not raw generation, become the scarce and valuable resources.

Wix is the most instructive competitor because OpenAI partners with it precisely to do what Sites cannot. Wix generated $1.99 billion in 2025 revenue and serves 250 million-plus registered users - GlobeNewswire. In June 2025 it acquired the vibe-coding startup Base44 for about $80 million upfront plus earn-outs, and Base44 reportedly hit $100M ARR within nine months of the deal - Wix Press Room. Wix builds real public websites with backends, payments, scheduling, and fulfillment, capabilities Sites does not offer. Wix CEO Avishai Abrahami has argued that a true ChatGPT-native site and commerce platform is structurally near-impossible to build, requiring a database "probably a billion times more complicated than the most advanced Oracle database" - Wix Blog. The OpenAI-Wix partnership is the strongest evidence that even OpenAI does not see Sites as a Wix replacement.

The design-first builders round out the public-website side of the market and each has a distinct 2026 strategy. Webflow restructured its plans in May 2026 into Starter, Basic ($15/month), and Premium ($25/month) tiers, layering an AI Site Builder and AI Assistant onto its visual canvas - Webflow. Framer reached a $2 billion valuation in 2025, positions itself as "the website builder loved by designers," and starts at $10/month - TechCrunch. Squarespace, taken private by Permira for $7.2 billion in late 2024, leans on its Blueprint AI builder, which can generate a custom site in under four minutes, with plans from $16/month - Permira. These tools own the public marketing-site category that Sites, with its authenticated-only sharing, simply does not enter.

The AI labs themselves are the most direct philosophical competitors, because they are building the same "describe it and it appears" capability into their chat products. Anthropic offers Claude Artifacts for chat-to-app creation, launched interactive Claude Apps in January 2026, and has reportedly tested a full-stack app builder that leaked in April 2026, though as of June it remains unconfirmed and unreleased - Dataconomy. Anthropic's current flagship is Claude Opus 4.8, released May 28, 2026, which we benchmarked in detail in our Claude Opus 4.8 guide. Google counters with Gemini Canvas for building docs, apps, and slides, plus Antigravity 2.0, an agent-first development platform announced at I/O 2026, all powered by the Gemini 3.5 model family - Google. The structural threat all three labs pose to dedicated builders is identical: they price the build capability into a $20/month subscription rather than charging per app.

It is worth being precise about how far each lab's stack already reaches, because the depth varies more than the headlines suggest. Anthropic shipped Claude Apps in January 2026, interactive third-party tools that render inside a conversation through an MCP extension, with launch partners including Slack, Canva, Figma, and Box - TechCrunch. It followed with Claude Design for prototypes and marketing assets, rebuilt Claude Code around parallel sessions and scheduled Routines, and says users have created over half a billion Artifacts - Claude.com. Google's answer is broader still: beyond Gemini Canvas, the Gemini app generates self-contained interactive micro-apps inside chat, and Antigravity 2.0 adds a Managed Agents API that spins up an agent in an isolated Linux environment from a single call - Google Developers. The labs are not dabbling at the edges. They are building the full prompt-to-deployed-app pipeline into products that hundreds of millions of people already pay for, which is the existential question hanging over every standalone builder, OpenAI's Sites included.

Finally, Cursor (Anysphere) is worth noting as the tool OpenAI's Codex most directly competes against, even though it is aimed at professional developers rather than founders. Cursor reached roughly $2 billion in annualized revenue by February 2026 and was reportedly raising at a $50 billion valuation - TechCrunch. It scores oddly well in our table despite a low ease rating, because what it lacks in approachability it makes up in total ownership: the code is yours, in your repo, with no host that can unpublish it. That trade-off, control versus convenience, is the throughline of this entire market.

7. Pricing: What You Actually Pay

Pricing is where the "is Sites a website builder" question gets settled in dollars, because the cost structure tells you who the product is for. There is no standalone price for Sites. It is bundled into ChatGPT Business and Enterprise, which means the real cost of using it is the cost of those plans plus the Codex token usage the agent consumes while building. For a founder used to a $16/month Squarespace plan, that is a different universe.

The foundation is the ChatGPT plan ladder. Codex is technically included in every tier, but Sites is not, so the relevant entry point is Business - OpenAI Developers. The table below shows the current ladder, with the line that matters for Sites bolded.

On top of the seat cost sits Codex usage, which OpenAI shifted from per-message to API token-based pricing on April 2, 2026 - OpenAI Developers. Usage is metered in credits per million tokens and varies by model. The flagship GPT-5.5 running in Codex is billed at 125 credits per million input tokens and 750 credits per million output tokens, while the cheaper GPT-5.4 runs at roughly half that - OpenAI Developers. OpenAI's own published estimate is that Codex costs about $100 to $200 per developer per month on average, with heavy users paying more - Verdent. Building Sites consumes that same usage, so a realistic all-in cost for a Sites-active seat is the Business seat plus meaningful token spend on top.

The rate card does reward smart model choice, which is the main lever a cost-conscious team has. Lighter models cost dramatically less: GPT-5.4-mini runs at roughly 18.75 credits per million input tokens against 125 for the flagship, and Fast Mode doubles consumption, so an experienced operator can swing the bill by an order of magnitude simply by matching the model to the task - OpenAI Developers. For teams, OpenAI added a pay-as-you-go Codex-only seat with no fixed monthly fee and no rate limits, billed purely on token consumption, which is the structure that makes Sites usage easier to budget across departments - Verdent. The same April 2026 overhaul also cut the ChatGPT Business seat from $25 to $20 per user per month - WinBuzzer. None of this changes the fundamental shape of the cost, though. A founder still pays an enterprise seat plus variable AI spend to produce an internal artifact, where a dedicated builder charges one predictable consumer fee for a public, owned product.

Compare that to what the dedicated builders charge a founder for a comparable or better outcome, and the gap is obvious. The point is not that one number is universally cheaper, but that the builders bundle hosting, a public URL, and ownership into a single flat fee, while Sites layers usage-based AI spend on top of an enterprise seat.

To make that concrete, picture a two-person startup that wants to use Sites. They need at least the two-seat minimum on ChatGPT Business at $20 per seat, which is $40 a month before anyone builds anything, then Codex token usage on top, which OpenAI itself estimates at $100 to $200 per active builder per month - Verdent. Call it roughly $140 to $240 a month, and what they get is an internal tool hosted on OpenAI's servers that they cannot export. For the same money, Lovable at $25 or Replit at $20 hands them a public, owned app with code they keep. The math only tilts toward Sites when the builder is already inside a company paying for Business seats anyway, which is, once again, exactly who Sites was built for.

The structural takeaway is that Sites is priced like enterprise software because it is enterprise software. It assumes an organization with a Business contract and a budget for AI usage, building internal tools for employees who are already logged in. The dedicated builders are priced like consumer products because they serve founders and small teams who want a public artifact for a predictable monthly fee. If pricing is destiny, Sites was never aimed at the founder market at all. For a deeper breakdown of how these monthly costs compound into real annual spend, our AI app builders ranking includes a full pricing-comparison section.

8. Where Sites Wins and Where It Breaks

A fair guide has to be specific about where Sites genuinely shines, because for the right job it is excellent. The product's sweet spot is the internal work-product: a thing your team needs to look at together, that used to live in a spreadsheet or a slide deck and now wants to be interactive. A financial analyst can turn a static model into a live scenario planner that executives tweak in a browser instead of clicking through tabs - VentureBeat. An operations lead can spin up a project dashboard. A product team can build a launch hub with milestones, owners, and a readiness checklist - Kingy AI. For these jobs, the authenticated-by-default model is a feature, not a bug, and the speed from idea to shared URL is real.

The strongest competitive read is that Sites threatens internal-tooling and BI vendors, not website builders. One analysis called the broader update "OpenAI's most direct assault on horizontal SaaS," arguing it threatens the workflow layer that Tableau, Power BI, and internal BI teams occupy - The Next Web. If Codex can orchestrate the apps where the data already lives and publish the result as a Site, it competes with Retool, Notion, and internal dashboards far more than with Wix. That is the accurate competitive frame, and it is a large, valuable market in its own right.

The breakage points are equally important, and they cluster around ownership and trust. The most serious is vendor lock-in. Sites are hosted by OpenAI on Cloudflare Workers, there is no documented export-to-static-HTML or self-host path, and OpenAI reserves the right to unpublish or delete sites - Kingy AI. You are building on rented land. For an internal dashboard that is acceptable. For anything you want to own, it is disqualifying, and it is the single biggest reason Sites scores a 3 on ownership in the assessment table.

Reliability and compliance compound the concern, and these are not hypothetical. Every deployment is immediately live with no real staging, AI-generated apps still hallucinate on hard multi-step tasks (one report cites roughly 30% in difficult multi-turn settings), and Sites explicitly prohibits HIPAA-regulated health data and PCI card data, shifting all privacy and consent responsibility onto the builder - The Hacker News. On June 1, 2026, a malicious npm package posing as a Codex web UI with 29,000-plus weekly downloads was found stealing Codex authentication tokens, a reminder that the stack underneath Sites is itself a target - TechRadar Pro. The most honest summary came from a launch analysis that conceded the open question directly: "whether that orchestration layer can match the reliability, security, and auditability that enterprises require is the question the preview period will answer" - The Next Web.

The smaller print compounds the caution. Sites is explicitly a beta service that OpenAI flags as unsuitable for sensitive workflows, the platform may attach a "powered by ChatGPT" attribution to your pages, and reviewers note that uptime SLAs, data-residency controls, backup and recovery policies, and any static-HTML export path are all unspecified or absent at launch - Kingy AI. Output quality is prompt-dependent rather than automatic, too: good results require the builder to specify the audience, the core experience, the data handling, and the access scope, which is real work and not the one-line magic the marketing implies. For an internal dashboard refreshed weekly by the team that made it, these gaps are tolerable. For a customer-facing product or a regulated workflow, they are the difference between a prototype and a liability, and a founder should weigh them before treating Sites as anything more than a place to build internal tools quickly.

The security and lock-in risks are not unique to OpenAI, which is the broader lesson. We documented the underpriced security exposure across the entire AI-builder market in our AI website builders market map, and the pattern holds here: the more an AI assembles your app from opaque generated code on a host you do not control, the more these risks compound. The takeaway for a founder is not "avoid AI builders." It is "know who holds the keys to what you build."

9. What It Means for Founders Building a Real Company

Strip away the enterprise framing and the founder's question is simple: can OpenAI Sites help me build and run my business? The honest answer, today, is mostly no, and understanding why is more useful than any feature comparison. Sites builds work-products inside a company you already have, not a company you are trying to start. It assumes the org, the workspace, the logged-in colleagues, and the budget. A founder has none of those yet. What a founder needs is the opposite shape: a public front door, a product customers can reach without a workspace login, payment collection, and something they own outright.

This is the distinction between building internal tools and building a business, and it is the most important strategic frame in this guide. The vibe-coding builders, Lovable, Replit, Bolt, and v0, get a founder much closer because they produce public, ownable, full-stack products. But even they typically hand you an app, not an operating company, leaving you to assemble billing, email, an admin panel, and analytics yourself. The deeper a tool goes toward a complete operating business, the more of the founder's actual problem it solves, which is exactly the axis the assessment table measured as output scope.

This is the category where autonomous company builders sit, and it is worth naming the distinction precisely because the market conflates it with website building. A platform like Founden is positioned not to generate a page but to generate the whole company: from one natural-language description it produces a website, a customer app, an admin dashboard, billing, email, a database, and deployment, with the explicit promise that you own everything - Founden. Where a Codex Site is an internal artifact hosted on OpenAI's surface, the autonomous-company model treats the output as a real business you operate and keep. It belongs on the same comparison table as Sites precisely because it answers the question Sites does not: not "what can my team look at," but "what can I launch and run."

The founder who understands this stops asking "which tool builds the prettiest page" and starts asking three sharper questions. Do I own what I build, or am I renting it from a platform that can unpublish it? Does the output reach the public, or only my workspace? Does it cover the whole company, or just one slice? Run Sites through those three filters and its role becomes clear: a powerful internal tool for teams inside companies, not a launchpad for founders starting them. We laid out the full modern playbook for that launch, from validating demand to building without a technical co-founder, in our guide to starting a company in 2026.

The reason this matters so much is that building was never the hard part of starting a company, even before AI. The hard parts are demand, distribution, and survival. Most companies fail not because the product could not be built but because no one needed it or no one found it, which is why distribution, not code, is becoming the real moat. Founders increasingly invest in audience and channels before they invest in product, a landscape we mapped in our guide to founder communities. An AI that builds your internal dashboard does nothing for any of those problems. An AI that builds and runs your whole company, then frees you to spend your hours on customers and channels rather than on tooling, addresses the parts that actually decide whether a business lives. That is the lens through which a founder should read every launch in this space, this one included.

It is a worldview that the people building these tools share. Yuma Heymans (@yumahey), who founded the autonomous-company platform behind Founden and co-founded the AI recruiter HeroHunt.ai, has argued that you build an autonomous company "bottom up, from tools, processes and people, all the way up to autonomous organizational orchestration" - LinkedIn. That bottom-up framing is the useful lens for the whole field: Sites automates a tool, the vibe-coding builders automate a product, and the autonomous-company builders aim at the whole stack. Knowing which layer you are buying is the entire decision. With 594 million founders worldwide and the vast majority never raising a dollar, as we documented in our startup founders data guide, most builders need ownership and distribution far more than they need another internal dashboard.

For the non-technical founder specifically, the practical move is to invert the usual order of operations. The old advice was to find a technical co-founder, then build, then launch. When an agent can build the product, the scarce skills become choosing the right problem, reaching customers, and operating the business once it exists. A founder in 2026 should pick tools by asking which one removes the most operational burden, not which one writes the cleverest code. Sites removes a slice of internal busywork for teams that already exist. The platforms a founder genuinely needs remove the burden of running an entire company, which is a different and far larger promise, and it is the promise that matches the actual shape of the founder's problem.

10. The Future Outlook: Agents That Build Businesses

The most important thing about Sites is not Sites. It is the direction it points. OpenAI did not build a website feature; it built a publishing primitive for an agent, and the trajectory of every major player suggests the same destination: AI agents that do not just write code but run the systems that code produces. The interesting question for the next twelve months is not whether agents can build a dashboard, which they clearly can, but how much of an entire company they can build and operate end to end.

The market data says the runway is enormous. The AI-powered website builder market is valued at roughly $3.24 billion in 2026 and projected to reach $17.43 billion by 2035 at a 20.55% CAGR - Precedence Research. Gartner forecasts that 75% of new applications will be built with low-code or no-code approaches and that the low-code market alone will hit $44.5 billion - InfoWorld. By mid-2025, an estimated 35% of newly published websites were already AI-generated or AI-assisted - Hostinger. The shift from humans building software to agents building software is not a forecast; it is already the base rate.

The competitive dynamic that defines this future is the collision between AI labs and app-layer startups, and Sites is the opening shot. The labs (OpenAI, Anthropic, Google) are pushing build capability down into their chat subscriptions, betting that owning the model and the surface lets them absorb the application layer. The startups (Lovable, Replit, Vercel, Wix) are racing to own the customer relationship, the workflow, and the ownership promise before that happens. The same tension shows up in funding: AI-native builders are raising at valuations that only make sense if they capture a durable layer the labs cannot, a dynamic visible across the AI-thesis investors we tracked in our top US VCs guide.

For a founder, the strategic implication is liberating rather than threatening. When intelligence becomes a cheap, abundant input, the bottleneck shifts away from building and toward owning, distributing, and operating what gets built. The labs will keep making the building part nearly free; that is good news. The durable advantage moves to whoever controls the customer relationship and the business itself, which is precisely why ownership and distribution, not raw generation, are the dimensions that should drive a founder's tool choice. Distribution in particular is becoming the real moat, a theme we explored across channels in our best AI social media tools guide and one that no website builder, OpenAI's included, can hand you.

The adoption curve underneath the market numbers is what makes the direction feel inevitable. By 2026, 58% of small businesses use generative AI, up from 23% in 2023; 84% of developers use or plan to use AI coding tools; and Gartner projects 75% of new applications will be built with low-code or no-code approaches - Hostinger. Set that against the fact that roughly 27% of US small businesses still have no website at all, and the implication is a vast population of first-time builders who will meet the web through an AI that builds it for them - Zippia. Whether that AI hands them something they own, or something they merely rent, is the question this entire guide has been circling.

The shift also rewires how the web is found, which raises the stakes for who hosts your pages. Gartner predicts traditional search volume will fall 25% by 2026 as discovery moves to AI assistants, and that more than a third of web content will be produced for AI-powered search - Gartner. In that world a Site hosted behind a workspace login, invisible to crawlers and assistants alike, is structurally cut off from the new discovery layer, while an owned, public, indexable site is not. The choice of where your web presence lives stops being a hosting detail and becomes a distribution decision.

The endgame that Sites hints at is the autonomous company: a business where the website, the app, the billing, the content, and the operations are built and increasingly run by agents, with a human setting direction rather than assembling tools. OpenAI is building the work-product layer of that future. The vibe-coding builders are building the product layer. And the autonomous-company builders are aiming at the whole thing. The June 2 launch of Sites is best understood not as a website builder, but as one more piece of evidence that the question is no longer whether AI can build your company, but who owns it when it does.

Conclusion: A Decision Framework

OpenAI Sites is an excellent product aimed at a job most founders do not have. It turns Codex's output into a hosted, shareable, interactive work-product, and for teams inside companies that need live dashboards, planners, and internal tools, it is genuinely powerful and worth the Business plan. But it is in preview, locked to enterprise workspaces, hosted on OpenAI's surface with no export, and shared only behind a login. Calling it a Wix killer is a category error, and OpenAI's own partnership with Wix proves it.

Use this three-question framework to decide. If you need an internal tool for a team that is already inside a company, Sites or a platform like Retool fits. If you need a public, ownable product to put in front of customers, the vibe-coding builders like Lovable and Replit are the stronger choice, because you keep the code and reach the open web. And if you need an entire operating company, website, app, billing, and operations from one description, the autonomous-company builders such as Founden are built for exactly that, where the others stop at a single layer. Match the tool to the layer you actually need, and the noise around the launch resolves into a clear decision.

The deeper truth the launch reveals is that building software is becoming the easy part. What remains hard, and what will decide which founders win, is owning what you build, reaching the people who need it, and operating the business around it. Choose tools that give you ownership and distribution, not just generation, and you will be building on solid ground while the platform wars play out above you.

This guide reflects the AI build-tools landscape as of June 2026. OpenAI Sites is in preview, and pricing, features, and availability for it and every platform mentioned change frequently. Verify current details directly with each provider before relying on them. Revenue and user figures attributed to private companies are self-reported and not independently audited.