Does OpenAI still have a moat?
In our generative AI market deck, you will find everything you need to understand the market
SUMMARY
Yes, OpenAI still has a moat, but it now sits mainly in ChatGPT’s distribution, paid audience and growing workflow layer rather than in a lasting lead on model intelligence.
The model moat has thinned sharply. OpenAI can still return to the frontier and win important coding or agentic tasks, but Anthropic, Google and open-weight challengers are close enough to keep customers flexible.
ChatGPT is the defensive asset that rivals have not reproduced. Its enormous audience gives OpenAI a ready-made launch channel for agents, search, shopping, premium plans and work products without having to acquire users from scratch each time.
That audience is habitual, not fully locked in. Simple prompts move between assistants in seconds, while memory, project history, connected tools and recurring tasks create the more meaningful switching costs.
OpenAI’s developer platform remains huge, yet the raw API is becoming interchangeable. The harder moat will come from evaluations, permissions, persistent agents, deployment controls and other infrastructure that becomes embedded in how a team operates.
Anthropic has already shown that a rival can take valuable enterprise workloads without matching OpenAI’s consumer reach. OpenAI’s response is to spread beyond model access and own more of the completed work through products such as Codex and Presence.
Codex may matter more strategically than another benchmark win. A workspace that understands files, keeps project history, uses tools and produces editable work is much harder to replace than a chatbot that gives a slightly better answer.
ChatGPT’s scale creates a product-feedback advantage, not a simple proprietary-data monopoly. OpenAI can observe where users retry, abandon or correct tasks, while its business-data promises limit how directly enterprise deployments can feed the base model.
Google is the clearest structural threat because it controls the places where context already lives: Search, Android, Chrome, Gmail, Docs and Cloud. OpenAI has the stronger AI-first destination, but Google can make the assistant disappear into software people already use.
Compute is both a barrier and a liability. OpenAI can secure capacity few independent companies could assemble, but the commitments become dangerous if model prices fall faster than usage grows or expensive infrastructure sits underused.
The unresolved question is profitability. OpenAI has proved that demand can grow at extraordinary speed; it has not yet proved that revenue can outrun research, inference and infrastructure costs for long enough to fund the business without repeated giant financing rounds.
OpenAI’s moat therefore looks real but unfinished. It becomes durable only if ChatGPT turns its reach into recurring workflows that are difficult to replace and cheap enough to run at a sustained profit.
This market map, featured in our generative AI market deck, highlights top companies and startups in the generative AI market
What would an OpenAI moat actually mean now?
Today, OpenAI’s moat comes from the people, habits and paid work built around ChatGPT.
A benchmark lead does not last long enough to protect OpenAI by itself. Frontier rankings change after almost every major release, and different models win on coding, research, design, speed or price. A one-point lead may help sales for a quarter, but it gives customers little reason to stay for years.
The harder assets to copy sit above the model. ChatGPT has a huge everyday audience, tens of millions of subscribers, more than a million paying business customers and products that remember context, connect to company systems and complete longer jobs. Those advantages can survive a period when another laboratory has the better model.
We also have to ask whether OpenAI can make money from this lead. The company could remain hugely influential while struggling to cover training, inference, infrastructure, sales and partner payments. Its products eventually need to pay for all of that without another larger funding round every year.
Why does OpenAI’s moat look weaker lately?
OpenAI’s moat looks weaker because superior intelligence has become much easier to find elsewhere.
GPT-4 gave OpenAI a clear period when serious alternatives were scarce. These days, Anthropic can lead on difficult coding work, Google can offer a competitive model inside products people already use, and Chinese open-weight models can get surprisingly close at a much lower price.
The market has also learned how to avoid dependence on one provider. Developers route requests across several models. Large companies test multiple vendors before choosing a workload. Consumer users keep ChatGPT open while also trying Gemini, Claude or Perplexity. Being first is worth less when switching is this easy.
So far, OpenAI has lost exclusivity without losing relevance. Menlo Ventures found that its enterprise model share had nearly halved in two years, while OpenAI kept adding paying users and growing revenue quickly. The old moat has narrowed as a new one forms around distribution and workflow.
If you want more recent data on this point, please see our latest generative AI market report.
As this chart shows, and as featured in our generative AI market deck, search interest in LLMs has surged
Is OpenAI still ahead on AI models?
OpenAI is currently one of the world’s best AI model makers, with no lasting lead over the other front-runners.
Artificial Analysis now places GPT-5.6 Sol within roughly one point of Anthropic’s leading model on its broad intelligence index. OpenAI leads several agentic and coding evaluations, while Anthropic wins others. The frontier is tight, and the winner changes with the task.
GPT-5.6 also shows why OpenAI remains dangerous. On the independent Artificial Analysis Coding Agent Index, Sol reached 80, ahead of Claude Fable 5 at 77.2. In OpenAI’s published tests, the model completes some long professional tasks faster and at lower estimated cost than comparable Claude models. Vendor tests deserve caution, but GPT-5.6 still shows that OpenAI can respond fast when a rival pulls ahead.
OpenAI’s alumni have also built some of its fiercest rivals. Former employees helped create Anthropic, Safe Superintelligence, Perplexity and other companies, so the know-how does not stay inside one lab. OpenAI can still attract top researchers with vast compute, famous products and immediate access to hundreds of millions of users. It can keep returning to the frontier, but it cannot own it.
| Current comparison | OpenAI’s position | What it tells us |
|---|---|---|
| Artificial Analysis broad intelligence index | About one point behind the leader | The overall model gap is tiny |
| Coding Agent Index | GPT-5.6 Sol leads at 80 | OpenAI can still win important workloads |
| Cost and speed | Competitive on several frontier tasks | Efficiency is becoming as important as raw scores |
| Talent diffusion | Many major rivals employ OpenAI alumni | Research knowledge does not stay locked inside one company |
Are open models turning GPT into a commodity?
Open models are already pushing down the value of ordinary GPT-style intelligence, especially for buyers who care more about price and control than the absolute best answer.
Moonshot AI’s Kimi K3 recently reached third place on Artificial Analysis’s broad index and topped Arena’s frontend-coding ranking. The Associated Press reported that it could run at around half the cost of GPT-5.6 Sol in the comparison examined. A model that is slightly worse overall but much cheaper can be the better business choice for extraction, support, classification or high-volume coding.
Open models are spreading fastest among developers and startups, where teams can handle deployment complexity and switch quickly. Menlo Ventures found that Chinese open models were gaining visible traction across OpenRouter and vLLM usage even while their enterprise share remained small. Airbnb’s use of Qwen and Cursor’s use of an open model as an internal base show how these systems can quietly enter major products without becoming famous consumer brands.
Large companies remain more cautious. Menlo estimated that open models held only 11% of enterprise model usage in 2025, down from 19% the year before. Support, security, governance and reliability still favor closed providers. Open models are already a serious price threat; replacing full enterprise platforms will take longer.
If you want more recent data on this point, please see our latest generative AI market report.
This chart, featured in our generative AI market deck, shows annual VC investment in generative AI startups
Is ChatGPT still the default AI assistant?
ChatGPT is still the default AI assistant today, and this remains OpenAI’s strongest moat.
Company figures put ChatGPT at more than 900 million weekly active users and over 50 million consumer subscribers. Sensor Tower separately found that the mobile app passed one billion monthly active users faster than TikTok, YouTube or Instagram. Whichever metric we use, no standalone rival has built a larger habit around asking an AI for help.
That paid audience gives OpenAI a simple advantage. When it launches a new agent, search product, shopping feature or premium plan, it can put it inside an app those people already open. A rival has to convince them to download something else, rebuild context and pay again.
ChatGPT’s lead has still narrowed. Sensor Tower found that its share of unique AI-assistant users fell below 50% as Gemini and Claude grew. That sounds worse than it is: ChatGPT kept adding users while the whole category expanded faster. OpenAI has moved from near-monopoly to clear leader, which is weaker but still a very large position.
| Consumer measure | Current picture | Why it matters |
|---|---|---|
| ChatGPT weekly active users | More than 900 million | OpenAI reaches people at extraordinary frequency |
| ChatGPT consumer subscribers | More than 50 million | A huge audience already pays directly |
| ChatGPT mobile monthly users | More than 1 billion | The app has reached true mass-market scale |
| ChatGPT share of AI-assistant users | Below 50% | Rivals are growing without shrinking ChatGPT’s absolute audience |
Are ChatGPT users loyal or just used to it?
People keep coming back to ChatGPT, but most of them can still leave easily.
Moving a simple prompt takes seconds. People can ask ChatGPT, Claude and Gemini the same question, then keep whichever answer they prefer. There is no social graph to abandon and usually no complicated data migration. Power users have plenty of freedom to shop around.
Memory, projects and connected tools make switching more annoying over time. A user who has stored writing preferences, uploaded project files, connected email and trained ChatGPT through months of corrections loses more than a clean chat box by leaving. The friction becomes stronger when an agent can continue a project, remember decisions and take action across several apps.
OpenAI still faces one basic disadvantage: much of a person’s useful context already lives inside Google Workspace, Microsoft 365, Slack, GitHub or company databases. Google and Microsoft can reach that context from products they already control. OpenAI needs ChatGPT to become the place where the work actually happens, rather than an assistant that reads data stored elsewhere.
For now, the consumer moat comes from convenience and familiarity. That keeps millions returning, but it does not stop sophisticated users from maintaining several AI subscriptions.
This chart, featured in our generative AI market deck, looks at OpenAI’s strategy in generative AI
Is Google becoming OpenAI’s biggest threat?
Google is now OpenAI’s biggest threat because Gemini already sits inside products people use every day.
Alphabet recently reported 950 million monthly users for the Gemini app, up sharply from earlier periods. ChatGPT and Gemini publish different usage metrics, so the figures are not perfectly comparable, but the direction is obvious: Google has turned Gemini into a mass-market product rather than a side experiment.
Distribution gives Google several routes OpenAI lacks. Gemini can appear inside Android, Search, Chrome, Gmail, Docs and Cloud. A good-enough assistant placed in all of those products can win through convenience, even when ChatGPT remains slightly better for some tasks. Google can also subsidize AI with advertising, cloud and Workspace revenue while OpenAI has to make the AI product itself pay.
Google’s threat reaches far beyond consumer chat. It owns the cloud, custom chips, productivity suite, browser, mobile operating system and search engine. OpenAI has partners across many of those layers, but access through a partner rarely offers the same control as ownership.
OpenAI still has stronger AI-first branding and a product people deliberately seek out. Google’s advantage grows when users stop caring which model powers the answer. If assistants become background features inside existing software, OpenAI’s standalone distribution becomes less valuable.
If you want more recent data on this point, please see our latest generative AI market report.
Is OpenAI still winning developers?
OpenAI still commands huge developer usage, while switching between models has become normal.
OpenAI reports that its APIs process more than 15 billion tokens per minute. The platform now includes tool use, retrieval, realtime voice, multimodal generation, agent orchestration and evaluation features. A team can build much more than a basic chatbot without leaving OpenAI’s stack.
Years of early adoption also created hidden switching costs. Production systems contain prompts, tests, guardrails, monitoring rules and fine-tuning data shaped around OpenAI models. Changing the model may take one line of code; proving the whole product still works can take weeks.
Yet the industry has spent the past two years making switching easier. Model routers, common APIs and orchestration frameworks let teams compare providers or send each task to a different one. Menlo’s research shows enterprise spending spreading across OpenAI, Anthropic and Google rather than settling permanently on one platform.
To keep developers, OpenAI has to make the surrounding workflow harder to leave. Evals, permissions, persistent agents, deployment tools and enterprise controls create much more attachment than an API endpoint. The raw model API has become too easy to substitute.
This chart, featured in our generative AI market deck, shows annual funding in generative AI startups
Has Anthropic taken the enterprise AI lead?
Anthropic now leads several of the most valuable enterprise model workloads, especially coding.
Menlo Ventures estimated that Anthropic captured 40% of enterprise model spending in 2025, compared with 27% for OpenAI and 21% for Google. Coding drove much of that gap: Anthropic reached an estimated 54% share of enterprise coding-model spending, against 21% for OpenAI.
OpenAI still has much broader reach. More than one million businesses pay it directly, and more than nine million people use paid ChatGPT business plans. Enterprise products now produce over 40% of OpenAI’s revenue, even though Anthropic has deeper usage inside some technical teams.
Companies rarely choose one winner. A bank might use ChatGPT Enterprise for employees, Azure OpenAI for internal applications and Claude for coding. Put together, enterprise AI already looks like a portfolio of models and products.
OpenAI is now moving further into implementation. Its new Presence product packages agents, company policies, evaluations, system access and human escalation into a managed deployment. According to OpenAI, the version handling its own English-language phone support already resolves 75% of inbound issues without a person. Presence is early and limited, but it shows where the moat could move: from selling intelligence to owning a working business process.
| Enterprise measure | OpenAI | Anthropic | What it means |
|---|---|---|---|
| Estimated model-spend share | 27% | 40% | Anthropic leads the model layer in Menlo’s survey |
| Estimated coding-model share | 21% | 54% | Anthropic has a clear specialist advantage |
| Paying business customers | More than 1 million | No comparable public figure | OpenAI has much broader organizational reach |
| Paid business users | More than 9 million | No comparable public figure | ChatGPT has spread far beyond technical teams |
Can Codex become a moat of its own?
Codex is currently OpenAI’s best chance to build a stronger moat around completed work rather than model access.
OpenAI reports more than five million weekly Codex users, over six times the level seen when the desktop app launched. Developers remain the largest group, but non-developers already represent about 20% of users and are growing more than three times faster. People are using it for reports, spreadsheets, presentations, contracts, research and workflow automation as well as code.
A workspace like this is harder to replace than a chatbot answer. Once a team relies on it to understand files, run several tasks in parallel, use tools and produce editable work, moving away becomes a real project.
Codex still has to beat strong products. Claude Code created much of Anthropic’s enterprise momentum, while Cursor, GitHub and other development environments can route jobs across several models. OpenAI cannot rely on GPT being the preferred coding model forever.
Codex becomes genuinely sticky when skills, permissions, project history, team reviews and integrations pile up around it. Five million weekly users prove fast adoption, although durable lock-in will take more time to demonstrate.
This chart, featured in our generative AI market deck, compares the main business model options for generative AI SaaS platforms
Does ChatGPT’s scale give OpenAI a data moat?
ChatGPT’s scale lets OpenAI see how people actually use AI at a level few rivals can match.
Hundreds of millions of users expose ChatGPT to rare questions, confusing instructions, new languages and unusual workflows. Even when the conversations themselves are not used for training, OpenAI can learn from product behavior: where users retry, abandon a task, correct an answer or choose a different tool.
OpenAI’s privacy-protected studies cover millions of conversations and track how usage changes across age groups, countries, languages and types of work. Few competitors can watch general-purpose AI behavior across such a wide audience.
The enterprise side is more restricted. OpenAI does not train on Business, Enterprise or API customer data by default. That promise helps it win sensitive workloads, but it also means a large corporate deployment does not simply feed proprietary company knowledge back into the base model.
This scale helps OpenAI find common failures earlier, test features on a huge audience and see which workflows people repeat. Competitors can buy data or generate synthetic examples, but years of live product feedback are harder to reproduce.
Is OpenAI’s compute empire a moat or a trap?
OpenAI’s compute buildout may be its strongest moat and its biggest financial danger.
OpenAI reports that Stargate has already secured more than 10 gigawatts of planned US infrastructure, passing the original target ahead of schedule. It has also raised $122 billion in new funding and broadened its supplier base across Microsoft, Oracle, CoreWeave, Amazon, Nvidia and others. Very few independent companies could assemble that much capital, power and chip capacity.
The latest spending plan makes the risk impossible to ignore. The Wall Street Journal reported that OpenAI now expects roughly $750 billion of computing commitments through 2030, up from about $600 billion previously. That figure is several dozen times OpenAI’s latest annual revenue.
Large-scale infrastructure can lower costs when utilization stays high. OpenAI can negotiate better terms, design custom chips, keep popular products available and train models that smaller labs cannot afford. The economics turn ugly when capacity sits idle or model prices fall faster than usage grows.
OpenAI is paying an extraordinary amount to make sure it never runs out of capacity. That lets the company serve demand rivals might miss, while creating fixed obligations before anyone knows what customers will pay for intelligence several years from now.
If you want more recent data on this point, please see our latest generative AI market report.
This chart, featured in our generative AI market deck, illustrates how revenue is distributed across customer segments in the generative AI market
Does Microsoft protect OpenAI or hold it back?
Microsoft still helps OpenAI more than it constrains it.
Microsoft supplied the capital, cloud access and enterprise credibility that helped OpenAI turn a research lab into a global product company. Azure remains the primary cloud partner, and new OpenAI products generally ship there first.
The revised agreement gives OpenAI more room. Microsoft’s license now runs through 2032 on a non-exclusive basis, and OpenAI can serve products through other cloud providers. That flexibility explains the growing relationships with Oracle, Amazon and CoreWeave.
OpenAI still pays heavily for that help. It will keep sharing revenue with Microsoft through 2030, subject to a cap, while Microsoft remains a major shareholder. Microsoft can also sell OpenAI technology through Azure, GitHub and Microsoft 365, sometimes owning the customer relationship OpenAI would rather control itself.
Both companies still need the deal: OpenAI gets infrastructure and distribution, while Microsoft gets access to frontier models. OpenAI’s recent diversification shows that it no longer wants one partner sitting between its research, products and customers.
Can OpenAI turn its scale into durable profits?
OpenAI has proved that demand is enormous; it has not proved that the business can produce durable profits.
Audited figures reviewed by the Financial Times showed revenue rising from $3.7 billion in 2024 to $13.07 billion in 2025. That is more than threefold growth in one year. OpenAI has since said its monthly revenue run rate reached about $2 billion, with enterprise activities contributing over 40%.
Costs grew even faster. The same audited statements showed roughly $34 billion in total costs and expenses for 2025, including more than $19 billion in research and development. The operating loss was about $21 billion before other accounting effects.
OpenAI has several ways to improve the picture. Cheaper inference can lift margins. Agents can be priced around completed work instead of tokens. Advertising, commerce and enterprise software can earn more from the same user base. Presence and Codex already point toward products where customers pay for an outcome, which is easier to price than a raw model call.
Competition keeps the burden high. Anthropic and Google can match important capabilities, open models push prices down, and large buyers can split workloads across vendors. Revenue growth alone will not create the moat. OpenAI needs cost per useful task to fall faster than the price customers pay.
This chart, featured in our generative AI market deck, shows how AI video generation technology has evolved over time
Does OpenAI still have a moat?
Yes. OpenAI still has a real moat today, and ChatGPT is carrying most of it.
The model advantage has become thin. GPT-5.6 is excellent and sometimes leads, yet Anthropic, Google and open-weight challengers can match enough of its capability to keep customers flexible. Anyone judging OpenAI only by benchmark leadership will probably conclude that the moat has disappeared.
ChatGPT tells a different story. Its consumer reach, paid audience and business footprint remain unmatched, while Codex and Presence are pulling OpenAI deeper into recurring work. Competitors can copy model features faster than they can recreate that whole distribution system.
The moat still has weak spots. Users can keep several assistants. Developers can route between APIs. Google can bundle Gemini across an ecosystem OpenAI does not own. Anthropic has taken the lead in valuable enterprise coding workloads. OpenAI also carries infrastructure commitments large enough to overwhelm even spectacular revenue growth.
OpenAI’s moat has shifted from models to products. It should remain one of the leaders for years if ChatGPT becomes a place where people finish recurring work rather than a website they visit for isolated answers. Its durability now depends on making those workflows difficult to replace and profitable to run.
| Part of OpenAI’s position | Moat strength now | What it means |
|---|---|---|
| Frontier models | Moderate | Strong enough to compete, too easy for rivals to match |
| ChatGPT distribution | Strong | OpenAI’s clearest defensive asset |
| Consumer habit and memory | Moderate | Useful friction, but users can still multi-home easily |
| Enterprise reach | Strong | Broad adoption gives OpenAI many routes into companies |
| Enterprise model share | Moderate | Anthropic has taken important high-value workloads |
| Codex and workflow products | Rising | The best route toward deeper switching costs |
| Developer API | Moderate | Large scale, but increasingly interchangeable |
| Compute access | Strong but risky | Difficult to copy and expensive to carry |
| Pricing power | Weak to moderate | Competition keeps model prices under pressure |
| Economic moat | Unproven | Demand is real; durable margins are still the open question |
If you want more recent data on this point, please see our latest generative AI market report.
OUR METHODOLOGY
This analysis tests whether OpenAI still has a durable moat based on the evidence available today. We separated the question into model performance, consumer distribution, user behavior, enterprise adoption, developer usage, workflow integration, data advantages, infrastructure and economics.
We treated benchmark leadership as a temporary advantage rather than proof of a moat. Frontier rankings can change after one release, so we used independent model and coding evaluations to judge whether OpenAI remains competitive, not whether it owns an enduring technical lead.
For distribution, we looked at weekly users, mobile users, subscribers, paying businesses and paid business seats. ChatGPT and Gemini publish different usage measures, so we used those figures to understand scale and direction rather than present them as a perfectly matched market-share comparison.
We used Menlo Ventures’ enterprise estimates as the clearest available view of how model spending is splitting across OpenAI, Anthropic, Google and open models. Those figures are market estimates rather than audited vendor revenue, so they help compare competitive position more than exact company performance.
We assessed open models separately from full enterprise platforms. Price, control and developer adoption show how quickly ordinary model intelligence is becoming substitutable, while support, security, governance and reliability explain why the same models may spread more slowly inside large companies.
We treated memory, projects, integrations, evaluation systems, permissions and recurring agent workflows as stronger evidence of switching costs than a model API alone. The aim was to separate products that people can replace in minutes from operating layers that become embedded in a team’s work.
Compute commitments were considered both an access advantage and a financial obligation. Planned power, chips and cloud capacity can protect availability and lower unit costs at high utilization, but commitments are not the same as profitable demand or capital already converted into productive infrastructure.
For profitability, we focused on whether recurring revenue can cover inference, research, staff, sales, partner payments and OpenAI’s infrastructure obligations. Fast revenue growth is useful evidence of demand, but it does not establish an economic moat unless cost per useful task falls fast enough to produce durable operating cash flow.
We prioritized recent company disclosures, independent benchmarks, audited financial reporting and market research that added a checkable metric. Key sources include OpenAI’s GPT-5.6 launch material, Artificial Analysis’ Intelligence Index, Artificial Analysis’ Coding Agent Index, Menlo Ventures’ enterprise AI report, Sensor Tower’s AI app research, Alphabet investor reporting, Financial Times reporting on OpenAI’s audited figures, and The Wall Street Journal’s reporting on compute commitments.
Additional product and infrastructure sources include OpenAI on Codex, OpenAI on Presence, OpenAI’s business-data controls, OpenAI on Stargate, the OpenAI developer platform, OpenRouter usage and rankings, Moonshot AI’s Kimi product information, and Airbnb Engineering.
In our generative AI market deck, we identify pain points entrepreneurs should prioritize
Related blog posts
- Can OpenAI ever become profitable?
- Will Chinese open models beat OpenAI?
- Can OpenAI actually afford all this compute?
