a16z Top 100 AI Apps: ChatGPT Is 30x Claude, But the Real Race Is Just Starting

ChatGPT remains 30 times larger than Claude on web and 80 times larger on mobile. But that headline number obscures something more interesting: the three major AI platforms are diverging into fundamentally different businesses, and the next wave of consumer AI won’t look like a chatbot at all.

The Report at a Glance

This is the 6th edition of a16z’s Top 100 Gen AI Consumer Apps report, spanning three years of tracking. Olivia Moore, the report’s author and a16z partner, joins Anish Acharya to unpack the data. The big picture: ChatGPT has only reached 10% of the global population as weekly active users. Despite explosive growth since 2023, consumer AI is still in its early innings.

Three shifts define this edition: the platform war is becoming a specialization war, non-AI-native products like Canva and Notion now qualify for the list (Notion says half its new ARR comes from AI features), and AI is breaking out of the browser into desktop apps, AI browsers, and ambient tools.

Three Platforms, Three Strategies

The raw numbers tell one story: ChatGPT is 2.7x bigger than Gemini on web, 2.5x on mobile, and roughly 30x bigger than Claude on web, 80x on mobile. Sam Altman’s Super Bowl-era tweet was literally true: more people use ChatGPT’s free tier in Texas than Claude has globally.

But the platforms are diverging fast. Look at their app stores: ChatGPT and Claude both have 200+ apps, but only 11% overlap. Claude is doubling down on prosumer with premium data sources, research tools, science tools, financial data, plus integrations like Claude in Excel, PowerPoint, and Chrome. ChatGPT is going broad consumer: travel, nutrition, personal finance, marketplaces. Gemini lives in its own corner, where traction correlates almost perfectly with creative model releases (V3, Nano Banana 1, Pro, 2).

The monetization strategies map to these positioning choices. Claude is subscription-only, targeting users and companies who will pay premium. ChatGPT is pursuing the Google model: acquire everyone, convert some to subscriptions, monetize the rest through ads and eventually transaction cuts on bookings and purchases. Sam Altman has said explicitly: “We want to be the AI for everyone.”

The App Store Lock-In Thesis

The most forward-looking argument in the conversation is about compounding advantages. Moore sees three forces that could create real platform lock-in for ChatGPT:

Social lock-in: Group chats mean if your friends are on ChatGPT, you’d have to convince them all to switch. Network effects, not just product quality, start to matter.

Developer concentration: As app stores emerge, developers will build for the platform with the most users first, exactly like the iOS/Android dynamic. For consumer tools, that favors ChatGPT’s scale.

Authentication layer: The most interesting one. Altman has hinted at “login with ChatGPT” where your memory and tokens travel with you to third-party products. The developer doesn’t pay for inference (the user brings their own), ChatGPT gets identity lock-in, and the user gets personalization everywhere. If it works, it creates a reason to concentrate your digital identity on ChatGPT.

The open question: will people want their work and personal AI identities merged? Moore flags that segmenting memory across different personas (work self vs. personal self) is an unsolved infrastructure problem.

The Global AI Heat Map

For the first time, the report includes per-capita AI adoption data across countries. The results are counterintuitive.

Singapore ranks #1, followed by Hong Kong, UAE, and South Korea. The US is #20. Russia and China are below #50.

The Russia surprise: like China, Russia has built a parallel AI ecosystem out of necessity (sanctions). Products like GigaChat, Yandex AI, and DeepSeek dominate. Russia is DeepSeek’s #2 market after China. Combined ChatGPT and Gemini usage in Russia is only 15%, the lowest of any country.

Why the US lags on per capita: the top-ranking countries have workforces skewed heavily toward tech-first, white-collar, high-skill jobs. The US has a giant chunk of employment in retail, transportation, and other sectors AI hasn’t touched yet. But culture matters too. An Edelman survey found US trust in AI at 32%, while top-adopting countries run 50-70%. China’s favorability hits 80%.

“If you’re in the US, you have probably internalized this ongoing angst and questioning around my job, or AI is terrible for artists, or all of these other things that make people pick up or not pick up AI.”

Moore expects more geo-specific AI products to emerge, especially as creative tools diverge along cultural lines, just as movies differ radically between India, China, and the US.

Creative Tools: From Hallucination Advantage to Specialization

The first big generative AI product was actually Midjourney, not ChatGPT. Early editions of the Top 100 were dominated by creative tools, which benefited from hallucination: surprising, beautiful, original outputs were features, not bugs.

That’s shifting. Commodity image generation (memes, basic marketing, infographics) has been absorbed by ChatGPT and Gemini’s core models. Standalone image generators that survive on the list, like Ideogram and Midjourney, are either aesthetically opinionated or offer sophisticated workflows you can’t get from a chatbot.

Music, voice, and video are different. The big model companies have invested less here, allowing specialists to break out: Suno for music, ElevenLabs for voice, both now consistently in the top 15-20. Video is the most contested space, and the Chinese models (particularly Kance 2) are “head and shoulders above” US offerings because they can train on any data. This benefits aggregator platforms like Krea where users can switch between models.

Sora’s Social Experiment

Sora’s launch was massive: #1 on the US App Store for 20 consecutive days (roughly 150,000+ daily downloads to hold that spot), hit 1 million users faster than ChatGPT itself, and still has 3 million DAUs. But new downloads have dropped from 6 million/month at peak to 1.5 million.

What worked: the video model quality and the “cameos” concept, where real people grant their likeness for AI-generated videos. Jake Paul went viral as the first big celebrity to lean in.

What didn’t: content was exportable. People took their Sora videos to TikTok, Instagram Reels, and YouTube, where AI content competed against the best human-made content. The overall feed experience on those platforms was better because it combined both.

“I don’t think we’ve seen a social product yet succeed that’s entirely AI content. The emotional stakes just feel lower in some ways.”

The potential path forward: Sora has signed deals with Disney and other media companies. If it becomes the only place to make licensed fan videos of beloved characters, that’s a defensible niche. But a massive AI-native social network hasn’t emerged yet.

OpenClaw, Manus, and the Agent Explosion

OpenClaw is the conversation’s most striking data point. It wasn’t on the Top 100 rankings (data cutoff was January), but February data would have placed it at #30 on web, a massive debut. It’s now #1 in GitHub stars of all time, surpassing React and Linux.

But the growth story has a nuance: usage keeps accelerating in the technical community, while visits to the signup page have plateaued since early February. OpenClaw hasn’t fully “escaped containment” to non-technical users. Moore expects OpenAI (which acquired it) to productize it for mainstream consumers, and notes that OpenClaw’s architecture has inspired a wave of startup pitches: “How many pitches do we take a day where the founder says ‘I want to be OpenClaw for X’?”

The multi-model advantage matters: OpenClaw works across all models, which could be diluted if OpenAI locks it to a single provider. Moore notes they’ve kept it multi-model for now.

Manus is positioned as the consumer-grade OpenClaw. It made the web list and was acquired by Meta for $2 billion+. The ramp from zero to $200 million ARR in six to nine months is “best-in-class.” It was the first consumer-grade agent that could actually operate autonomously across products and platforms, connecting email, browsing the web, making slides and spreadsheets, at a time when ChatGPT Operator and Google’s Project Mariner weren’t reliable.

Moore’s broader take on horizontal consumer AI agents: once everyone has agentic capability (driven by the core models improving), very horizontal products may be better off with the distribution of a Meta or Google than fighting as a standalone startup. Vertical agents have more defensibility.

The Next Six Months: Voice, Memory, Agents

Three forces Moore sees reshaping the landscape:

Voice is the most information-dense input medium. Voice dictation has gone from engineers to broader tech workers in six months. Meeting recording and transcription by AI is now nearly the norm. Moore expects voice-first AI tools (dictation, voice pins, ambient assistants) to hit mainstream consumers within 6-9 months.

Memory will become a core competitive advantage. ChatGPT and Claude are already good at it, and Google has launched “personal intelligence” pulling from Docs, Gmail, and Calendar. The unsolved problem: memory crossing contexts inappropriately (your AI knowing work things in a personal context).

“Any product that you start to use 2 years from now, if it doesn’t immediately feel like it knows you, it will feel broken. The concept of onboarding to a product should not be something that exists in a couple years.”

Agents will become invisible. Moore compares it to “internet companies” in the 1990s: eventually every tech company was a dot-com, and eventually every AI company will be agentic. Teenagers won’t think of them as agents, but agents will unlock consumer use cases (finance, healthcare, travel, complex shopping) that required too much cross-system data gathering to work before.

Some Thoughts

This conversation is valuable less for any single revelation than for the comprehensive view it offers of how the consumer AI market is actually segmenting. A few observations worth sitting with:

• The 11% app store overlap between ChatGPT and Claude suggests the platform war may resolve not through one winner but through specialization. Claude is becoming the Bloomberg Terminal of AI; ChatGPT is becoming the Google of AI. Both can be massive businesses.
• The per-capita adoption data quietly undermines the narrative that the US leads in AI. American cultural anxiety about AI, measured at 32% trust, is a real headwind for adoption, even as American companies build the best models.
• OpenClaw’s plateau in non-technical signups is the most important signal for anyone building consumer agents. The technology works; the product packaging for normal people doesn’t exist yet.
• Moore’s prediction that “onboarding should not exist in a couple years” is the most provocative claim. If memory and identity become truly portable, the switching costs in software fundamentally change, potentially destroying the current SaaS model’s reliance on setup friction as retention.