Epoch AI and Ipsos surveys show that 80% of Claude’s weekly active users in the United States come from families with an annual income of more than $100,000. AI assistants began to be stratified by price, entrance, and work scenarios, with high-income users taking the lead in entering higher-level AI services. A U.S. national survey revealed the differences in user profiles between major flagship models.
79.8% of U.S. adults who used Claude in the past week were from households with an annual income of $100,000 or more.
This ratio is higher than Microsoft Copilot's 63.7%, ChatGPT's 60.3%, Grok's 56.2%, Google Gemini's 55.9%, and much higher than Meta AI's 36.5%.
For reference, Epoch AI used U.S. Census data to estimate that about 50% of U.S. adults live in households with an annual income of more than $100,000.
The gap at the low-income end is equally stark.
Among Claude’s weekly active users, households with an annual income of less than $50,000 account for approximately 6.4%;
Meta AI corresponds to 32.1%.
About 24% of U.S. adults live in households with an annual income of less than $50,000.
The survey was conducted by Epoch AI in partnership with Ipsos, using Ipsos’ KnowledgePanel, with a sample of U.S. adults drawn from address-based probability sampling.
The first wave of the survey was conducted from March 3 to 6, 2026, with a sample size of 2,021 people, and the overall error at the 95% confidence level was plus or minus 2.2 percentage points; Epoch’s subsequent analysis of income distribution combined the three rounds of survey data in March and April.
A more straightforward statement is that Claude has shown a high income concentration in the United States.
It has not yet become the default entrance for the public, but it is becoming a tool used more frequently and intensively by a certain group of people.
Claude didn't win at scale.
Income portraits can easily create misinterpretations.
Claude's users are wealthier, but Claude's user base is still small.
In this round of Ipsos national survey, 31% of American adults have used ChatGPT in the past week, 21% of Google Gemini, 11% of Microsoft Copilot, 8% of Meta AI, 5% of Grok, and only 3% of Claude.
Another 49% of U.S. adults said they had not used any AI services in the past week.
The Decoder quoted Epoch AI as adding another layer of data. Among people with an annual income of more than $100,000, ChatGPT’s reach rate is still 37%, Gemini is 24%, Copilot is 14%, and Claude is only 6%; at the same time, 44% of high-income people have not used AI services in the past week.
So Claude's situation is more like high income concentration but low absolute coverage.
It appears to be more elitist in a small pool, and is far from approaching ChatGPT’s default status in the entire US AI market.
This also explains why this set of data is newsworthy.
In the past two years, AI companies have mostly used monthly activity, downloads, and calls to tell stories.
Now, user structure starts to matter more than user size.
Differences begin to emerge among who is dabbling, who is paying, and who is plugging AI into their workflows.
Why Claude
Claude's high-income profile is difficult to explain solely by price.
Anthropic's Claude Pro is $20 per month, Max 5x is $125 per month, and Max 20x is $250 per month, with Max aimed at users who need higher usage, less disruption, priority access to new models and features, and includes Claude Code.
OpenAI's ChatGPT also has $20 Plus and higher-end Pro plans, and personal high-usage AI products are entering the $100 and $200 price ranges.
The difference is product mentality.
ChatGPT is more like a universal entrance.
Gemini is bundled with Google search, Gmail, Docs and other scenarios.
Copilot is deeply connected to Microsoft 365, Word, Excel, Teams, and Edge.
Meta AI goes directly to WhatsApp, Instagram, Facebook, and Messenger.
Typical usage scenarios for Claude are more active access, long text processing, coding, complex writing, and professional tasks.
It requires users to know why they are opening it, and it makes it easier to attract people who are already willing to pay for efficiency.
Ipsos’ paid subscription data also illustrates this point.
For ChatGPT paid subscriptions, 4% of respondents said they paid for it themselves, and 3% said they were paid by their employer or school; for Claude, the corresponding proportion was 1%; for Copilot, 5% paid for themselves and 10% were paid by their employer or school.
Claude's paying audience is very narrow, but these people are more likely to be high-willing and high-intensity users.
That's where Anthropic is now, smaller and potentially more valuable per user.
It does not rely on social products to spread like Meta AI, nor does it rely on search portals to distribute like Google.
Claude requires active user choice.
But active choice itself is the threshold.
Meta AI is on the other side
Meta AI is at the other end of the table.
Among its weekly active users, only 36.5% are from households with an annual income of more than $100,000, and 32.1% are from households with an annual income of less than $50,000.
Among this group of mainstream AI assistants, it’s the closest to mass market.
The reason is not complicated.
An Ipsos survey shows that among people who have used Meta AI, 55% came into contact with it through the built-in functions of WhatsApp, Instagram, Facebook or Messenger, 40% saw AI-generated summaries or answers when searching on Facebook or Instagram, and only 21% entered questions into meta.ai or Meta AI applications.
The entrance determines the user.
Meta AI is put into social networks, Gemini is put into search, and Copilot is put into office software.
Claude relies more on users entering the product with a clear mission.
This has completely different business consequences.
Meta AI can reach a wider range of people, but user intentions are more dispersed, and many interactions may just be casual questions.
Claude has fewer users, but they are more likely to come in with a work problem, and their needs are clearer and easier to convert into subscriptions, API calls, or enterprise purchases.
The AI market is repeating the old story of consuming the Internet and productivity software, with huge traffic entrances on one side and high ARPU (Average Revenue Per User) on the other.,average revenue per user) tool.
The former is responsible for coverage and the latter is responsible for collecting money.
watershed in usage intensity
The more critical stratification occurs not only in whether AI has been used, but also in how to use AI.
The Ipsos survey shows that among people who have used AI services in the past week, 34% only used it for one day, 49% used it for 2 to 5 days, and 16% used it almost every day.
On the day of heaviest use, 62% only handled one or two quick tasks, 32% used it multiple times, and only 6% reported heavy use or heavy reliance on AI that day.
This shows that the penetration rate of AI in the United States seems to be quite high, but most use is still very light.
A large number of users just use AI as a search box, rewriter, and temporary question and answer machine; a small number of users start to use it as a work interface.
The same goes for work situations.
Among employed AI users, 46% mainly use it for personal matters, 26% mainly use it for work, and 25% use it for work and personal matters;
Of those who use AI at work, 33% use services paid for or provided by their employer, 50% use personal subscriptions or free services, and 11% use both.
Putting this set of figures next to Claude's income profile will give a clearer judgment. The next round of competition in the AI industry is likely to revolve around high-intensity users.
High-intensity users don’t just ask about the weather, write emails, or summarize web pages.
They will integrate AI into coding, contracts, sales, research, placement, procurement, customer service and data analysis.
The greater the gap in model capabilities, the greater the gap in results brought by tools.
Anthropic’s own Project Deal experiment provides an interesting sideshow.
In the experiment, different Claude models agent employees to buy and sell real items in the internal market.
The stronger Opus model paid an average of $2.68 more for the same item as a seller and $2.45 less as a buyer.
When Opus sellers faced Haiku buyers, the average transaction price was $24.18, higher than the $18.63 for Opus-to-Opus transactions.
What's more subtle is that users at a disadvantage do not clearly feel that they are at a disadvantage.
Such experiments are small-scale and occur within the company and cannot be directly generalized to the entire business world. But the direction is clear.
When AI begins to negotiate, purchase, write code, and do research on behalf of people, model capabilities will become a new means of production.
Whoever uses a stronger model and puts the model into the workflow earlier is likely to get better results.
The gap will be hidden in every small decision.
Signals for OpenAI, Anthropic and advertisers
The survey points to different issues for several companies.
For Anthropic, Claude's high-income profile is both a boon and a pressure.
The good thing is that it proves that Claude has attracted a group of users who are more likely to pay and more likely to use it intensively.
The stress is that Claude is still too small.
A weekly usage rate of 3% cannot support a mass platform narrative.
Anthropic needs to continue to improve high-end user value while finding distribution methods with lower barriers to entry.
For OpenAI, ChatGPT remains the default entry.
It leads both in overall weekly usage and in reaching high-income groups.
The real challenge is to transform the advantages of scale into higher-order workflow locking to prevent high-value users from flowing to Claude in scenarios such as coding, research, and long documents.
For Meta, Meta AI has a higher proportion of low-income users, which shows that it is close to the public.
Its commercialization path may not rely on high-priced subscriptions like Claude, but more like an extension of advertising, recommendations, search, and content consumption portals.
For advertisers and enterprise software companies, this is an early adopter map.
Behind different AI assistants, there may be different purchasing power, different task types, and different conversion paths.
ChatGPT represents the largest default entrance, Claude represents higher income and more professional usage tendencies, Copilot represents enterprise distribution in office software, Gemini represents search and Google ecosystem, and Meta AI represents mass reach in social networks.
Of course, the data must also be viewed with caution.
Epoch AI stated in the income analysis that the sample size of Claude is 201 and Grok is 221. Compared with ChatGPT and Gemini, the sample size is smaller and the confidence interval is wider;
The survey is a cross-sectional sample and usage is self-reported, subject to recall error and misclassification.
But the trend is clear enough.
AI will not only be sorted by model ability, but will also be re-queued by user class, entry location, willingness to pay and work intensity.
Over the past two years, the industry has been paying attention to whose model is stronger.
The more valuable questions then become, who is using the strongest model, who is paying for it, and who is making it part of their daily routine.