Only 4.5% of Americans Pay for AI, and the Top 1% Spend $903 a Month: a16z Top 100

Consumer AI is used by almost everyone and paid for by almost no one. That is the headline of Andreessen Horowitz's seventh Top 100 Gen AI Consumer Apps report, published October 5, 2026, and the first edition to add a ranking based on what US consumers actually spend. In August, just 4.5% of eligible US consumers held a paid personal subscription to ChatGPT, Gemini or Claude, and the top 1% of payers averaged $903 a month. The Decoder summarized the findings on October 10.

For anyone choosing a plan, building a consumer AI product or trying to understand why every lab keeps adding $100-plus tiers, the numbers are worth reading closely. They also come with caveats that most headlines drop, so this post separates what a16z measured from what it inferred.

TL;DR: the numbers in one table

What is new in the seventh edition?

The first six lists ranked products by monthly web visits (Similarweb) and mobile active users (Sensor Tower). That approach hides a lot, as a16z admits: "traffic data alone increasingly conceals the real trends in consumer AI." This edition keeps both traffic rankings, with a top 50 for web and a top 50 for mobile, and adds a third list ranking vendors by observed spending on US consumer cards, supplied by YipitData.

Two other changes matter. Only 11 products debuted, the fewest of any edition, which a16z reads as the top of the market hardening. And products primarily designed for NSFW use are now excluded from the rankings; a16z says they would have made up more than 20% of the web traffic list and are being treated as a separate category.

The spending list surfaces businesses the traffic lists miss. Of the top 50 vendors by spend, 29 appear on neither the web nor the mobile traffic list, including desktop apps and agents that live inside messaging platforms. Only seven companies make all three lists: ChatGPT, Claude, Suno, Perplexity, Photoroom, Canva and Notion. ChatGPT is first on each.

Wide but shallow: why only 4.5% pay

Subscription pricing tiers shown as ladder boxes, illustrating how consumer AI plans stack up from free to premium

The 4.5% figure covers the three general assistants only, and it roughly doubled in a year. That is fast growth from a small base. a16z's reading is that free tiers are good enough for casual use, while paid plans offer more features but demand a steeper learning curve, so many free users never get past the surface.

Two supporting facts from the report: only 13% of people who pay for one AI product pay for any other AI tool, and only about half of active AI users say they use it daily. In other words, the typical paying customer is loyal to a single product rather than stacking subscriptions. That makes the plan you choose, and how often it changes its limits, a bigger deal than it looks. We track those changes in posts like Anthropic's monthly API credits for Max and Team plans and Gemini's free and AI Plus plan changes.

Spending follows a power law

Rising green candlestick bars representing the concentration of AI spending in a small group of heavy payers

The most striking section of the report is about concentration. According to a16z:

• The top 1% of payers account for 19.5% of all observed consumer AI spend, more than the bottom 50% combined (16.6%).
• That top 1% pays $903 a month on average on consumer cards, and a16z adds that work usage may be even higher.
• Their spend is up about 80% over 18 months, while the median payer's $25 a month has barely moved.
• The top 10% of spenders account for roughly half of observed spending.

The Decoder rounds this to "about $900"; the exact figure in a16z's own text is $903.

Who are these heavy payers? a16z says they look less like mainstream consumers and more like prosumers: people buying software to build, create and get work done. They over-index on automation and building tools such as n8n, fal, Manus and Nous Research's Hermes Agent, and on creative tools such as Higgsfield, Figma and HeyGen. That matches the picture in our coverage of Claude Cowork use cases and the shift toward agent-style products, where one person can burn through far more compute than a casual chatter.

Expect this to shape pricing. If a sliver of customers generates a fifth of revenue, labs will keep inventing higher tiers and usage-based add-ons for them, while keeping the cheap plans as a funnel.

ChatGPT, Claude and Gemini: three different games

Coding assistant laptop with a green spark, representing prosumer coding and productivity tools that drive Claude subscriptions

At the top of the traffic ranking nothing changed: ChatGPT leads Gemini roughly 2x on August web visits and Claude about 6x. On mobile the gap is wider, with ChatGPT at 2.5x Gemini and 14x Claude in monthly active users. In the YipitData panel, ChatGPT has about three times as many US consumer paid subscribers as Claude or Gemini.

The biggest story is Claude as the clear number three. It did not appear on the first web list in September 2023 and has since passed DeepSeek and Perplexity on traffic. a16z also says Claude passed Gemini in US consumer subscribers earlier this year, helped by Cowork in January, Claude Design in April, the Fable 5 model in June and a burst of public attention around Anthropic's February dispute with the Pentagon, when Claude hit number one among free US mobile apps for the first time. After Google moved legacy paid users onto its AI plan, which inflated Gemini's subscriber count, the two are about even.

Anthropic also monetizes harder. a16z says 7.3% of Claude's consumer payers are on the Max plan, which starts at $100 a month, versus 1.3% at Google and 1.1% at ChatGPT on their corresponding $100 tiers. Anthropic, unlike OpenAI and Google, has said it will not run ads in Claude, so subscriptions are its consumer model.

The momentum shifted in midsummer, per a16z. YipitData's desktop panel recorded lower average daily Claude sessions in July and August, and the e-receipt panel showed gross additions slowing and churn rising. ChatGPT re-accelerated with the July launch of GPT-5.6 Sol, Terra and Luna and the debut of ChatGPT Work. Only 8% of US ChatGPT subscribers also subscribe to Claude. For a hands-on view of that product, see our ChatGPT Work vs Codex guide.

Note that Gemini is missing from the spending ranking: as The Decoder notes, its revenue cannot be separated from Google's other subscriptions.

Personal agents: the next monetization test

Six months ago a16z argued agents like OpenClaw were not usable by most consumers. Its new report says several startup agents now claim hundreds of thousands of users, including Instinct, Tomo, Poke, Lindy and Town, while incumbents answered with Meta's Muse, OpenAI's Dots and xAI's Grok Bot. We compared the big three in OpenAI Dots vs Grok Bot vs Meta Muse.

The numbers a16z cites are founder claims or reports, not audited data:

• Instinct founder Noah Shinn says 40% of users link a personal credit card within three weeks and then spend an average of $1,300 a month through the service, and claims more than $1 billion in annualized transaction volume, half of it travel.
• Muse reportedly reached 250,000 daily actives in its first week and crossed 5 million downloads in under a month, but Meta's Threads logged more than 15 million downloads in its first 22 days across the same two markets.
• Platforms are choosing sides: Amazon shut Muse out in under two weeks, as we covered in Amazon blocks Meta Muse's shopping agent, while Shopify, Instacart, OpenTable, Expedia, Ticketmaster and Plaid signed official integrations. See also Muse's transaction layer and our read on Muse's reported daily-user plateau.

Why does this matter for the business-model question? Agents that complete purchases can earn affiliate fees or a take rate instead of a subscription, which could make a free or near-unlimited assistant viable. Both Instinct and Muse have described plans to earn transaction fees, but a16z notes neither does so yet, and ChatGPT currently does not charge for purchases that start in the product. For the security side of agents holding cards, see Instinct and 1Password's account vaults.

The business-model problem

a16z's final argument is that consumer AI "needs a new business model, or possibly an old one." Of the 44 AI-native products in its top web ranking, all have a monetization model live. Subscriptions are offered by 84%, usage charges or extra credits by 64%, ads by 14% and transaction or platform fees by 2%. Products can use more than one.

That inverts the previous internet era, where in 2025 advertising generated 97.6% of Meta's revenue and 73.2% of Alphabet's. a16z's explanation is cost: when users were nearly free to serve, companies could burn cash to build a large audience first, but model costs make forgoing early revenue hard. It expects this to ease as open models spread and builders route simple tasks to cheaper models.

Alternatives are appearing. OpenAI said ChatGPT advertising reached a $1 billion annualized run rate in August, on 1.2 billion weekly active users, and an estimated 50 to 60% of US physicians use OpenEvidence, which now partly monetizes via ads, per a16z.

What this means for what you build or pay

• If you pay for AI, you are in a small minority, and likely a heavy user. The $25 median plan is the norm; $100-plus tiers are priced for people who run agents and long coding sessions all day.
• If you build a consumer AI product, the data favors differentiation. a16z's examples of startups still winning: those with a proprietary model (Suno is number 19 on web and number 7 on spend), a multi-model experience (Cursor, Lovable, Replit), a specific audience (OpenEvidence, Venice) or an experience that sidesteps the blank chat box (Plaud, a hardware plus subscription notetaker that debuted at number 16 on spend).
• If you build agents, plan for platform pushback and a pricing model that is not a flat seat.
• If you read the headlines, remember the panel limits. The figures are US-only and skip employer-paid seats, API spend and anyone who pays outside a tracked card.

Limits of the data

a16z itself says the spending data comes from panels rather than a census and is not total company revenue. YipitData tracks e-receipts and cards, so purchases through app stores or corporate billing may be under-represented. The Instinct and Muse figures are company or press claims. And a16z is an investor in the AI ecosystem with an interest in how the market is framed, so treat the narrative sections as analysis and the panel numbers as the part that can be checked against future editions.

Related reading

• Anthropic monthly API credits for Max and Team plans
• Gemini free and AI Plus plan changes
• OpenAI Dots vs Grok Bot vs Meta Muse
• Amazon blocks Meta Muse's shopping agent
• Meta Muse's transaction layer
• ChatGPT Work vs Codex guide
• Top 10 Claude Cowork use cases

Figures are as published by a16z on October 5, 2026 and may be revised in later editions.