CHAPTERS
What’s new in a16z’s Top 100 AI Apps (and why it still feels early)
Anish and Olivia set the stage for the sixth edition of the Top 100 AI Apps report, highlighting how fast the landscape is shifting while overall adoption remains surprisingly low. Olivia frames the biggest changes: an intensifying consumer race, legacy apps becoming “AI-majority,” and AI spreading beyond the prompt box into new surfaces.
- •ChatGPT remains the dominant global AI product, yet only ~10% of the world uses it weekly
- •The consumer/prosumer race is heating up across ChatGPT, Gemini, and Claude
- •Report now includes non-AI-native products that are now heavily AI-driven (e.g., Canva, Notion, Freepik)
- •AI is expanding into browsers, desktop apps, and embedded productivity contexts (Excel/PowerPoint/Chrome)
Foundation models in practice: where ChatGPT, Claude, and Gemini specialize
Olivia breaks down real-world usage share and how each major assistant is carving out a differentiated position. While ChatGPT leads by a wide margin, Claude and Gemini are building distinct ecosystems and use-case gravity, creating a multi-tool reality for users.
- •ChatGPT leads usage massively: ~2.7× Gemini on web, ~2.5× on mobile; far ahead of Claude
- •Claude is positioning toward prosumer/workflows (e.g., Claude Code, Office integrations)
- •App ecosystems differ: 200+ apps each on Claude/ChatGPT but only ~11% overlap
- •Gemini’s traction is tightly linked to creative model releases and Google product integration
The AI ‘app store’ dynamic: monetization paths and compounding distribution
They explore the emerging app directory/app store concept and how business models differ across platforms. ChatGPT is positioned as a broad consumer gateway with multiple monetization levers, while Claude leans into subscription and high-ACV/professional integrations.
- •ChatGPT’s “AI for everyone” strategy targets broad adoption first, monetization later
- •Claude is clearer on subscription-first monetization, skewing toward users/companies who can pay
- •ChatGPT’s app strategy resembles Google’s: future monetization via ads and transactions (e.g., travel booking)
- •Distribution and developer focus may compound toward the platform with the most users
Lock-in and ‘context compounding’: identity, network effects, and log-in layers
Olivia explains why memory and context may become less portable over time, increasing platform lock-in. They discuss network effects (group chats), developer incentives, and a potential “log in with ChatGPT” layer that could let users carry identity, memory, and inference to third-party apps—along with privacy tensions between work and personal use.
- •Lock-in can grow via social features (e.g., group chats) and classic network effects
- •Developers may prioritize building/shipping first on the platform with the biggest user base
- •Potential ‘log in with ChatGPT’ could let users bring memory/tokens (and reduce dev inference cost)
- •Open question: how to separate work vs personal identities/memory (‘don’t cross the streams’)
Google’s Gemini comeback: DeepMind-led creativity vs inertia in legacy products
They assess Google’s shift from Bard-era skepticism to a more confident, creative, multimodal push. Olivia and Anish argue Google’s best breakthroughs come from greenfield efforts (like NotebookLM), while Sheets/Docs face organizational and product inertia that slows radical change.
- •Google is cautious about baking AI into core products due to cannibalization and user switching costs
- •DeepMind-driven creative products are the standout accelerant (NotebookLM as an early signal)
- •Greenfield products move faster than legacy surfaces burdened by internal overhead
- •Google may rely on enterprise lock-in in the near term while selectively innovating
Global AI adoption: parallel ecosystems in China and Russia
Olivia shares new geographic analysis showing how restrictions and sanctions create alternate AI stacks. China and Russia emerge as major outliers with strong local ecosystems and comparatively low usage of Western assistants, shaping global competition and product rankings.
- •China has the lowest combined ChatGPT+Gemini usage; local tools dominate (e.g., Doubao, DeepSeek, Kimi)
- •Russia similarly relies on domestic/region-specific tools (e.g., GigaChat, Yandex) plus DeepSeek
- •Russia is DeepSeek’s #2 market after China
- •Outlier markets are large enough to affect global ‘Top 100’ standings over time
Per-capita heat map: who adopts AI fastest—and why the US is only #20
They discuss per-capita adoption across the top LLM products and what it reveals about workforce composition and cultural trust. Smaller, tech-forward economies top the list, while the US lags behind due to job mix and lower trust/optimism toward AI.
- •Top per-capita adopters: Singapore (#1), then Hong Kong, UAE, South Korea
- •US ranks ~#20 despite producing many leading AI products
- •Workforce composition matters: more tech/white-collar jobs correlate with higher usage
- •Trust varies widely: US trust cited ~32% vs 50–70%+ in many faster-adopting countries
Creative tools evolve: from standalone image generators to music/voice/video breakouts
Olivia traces how creative AI shifted from early dominance (Midjourney era) to a world where basic image generation is commoditized by the big assistants. Standalone winners now either have strong aesthetic/workflow differentiation or focus on categories where foundation model platforms invested less, like music and voice; video remains unsettled and multi-model.
- •Early creative tools benefited from ‘hallucination’ producing surprising, beautiful outputs
- •Standalone image generators are fewer as ChatGPT/Gemini handle commodity images well
- •Differentiated survivors: aesthetic opinionation or advanced workflows (e.g., Midjourney, Ideogram)
- •Breakout categories: music (Suno), voice (ElevenLabs); video is fragmented with strong Chinese models (e.g., “CDance2”)
- •Aggregators like Krea benefit if no single video model ‘rules them all’
Sora as a social experiment: explosive launch, exportable content, and status games
They analyze Sora’s launch metrics and why its social ambitions are harder than its creative-tool value. Sora scaled rapidly and maintains meaningful DAUs, but social retention is challenged because creators export the best content to TikTok/IG/YouTube where it competes with top human-made media; a fully AI-content social network still hasn’t emerged.
- •Sora hit #1 in the US App Store for ~20 consecutive days and reached 1M users faster than ChatGPT
- •Still meaningful usage: ~3M DAUs cited, but new downloads declined from peak
- •‘Cameos’ let people grant likeness for memeable videos; early viral moments (e.g., Jake Paul)
- •Exportability weakens in-app feed differentiation; AI-only feeds often feel lower-stakes emotionally
- •Potential niche: licensed creation via media deals (e.g., Disney) enabling authorized fan videos
The agents surge: OpenClaw, GitHub dominance, and what ‘consumer-grade’ looks like
Olivia recaps major agent momentum in the last couple months, focusing on OpenClaw’s technical adoption and Manus’s consumer breakthrough. OpenClaw becomes a developer phenomenon (GitHub stars) but hasn’t fully crossed into mainstream onboarding, while Manus proved autonomous cross-app workflows can work for consumers—though horizontal agent products may need big-tech distribution to win long term.
- •OpenClaw would have debuted around #30 on web if included; data cutoff excluded it from rankings
- •OpenClaw became #1 all-time by GitHub stars (surpassing React and Linux)
- •Technical adoption accelerates, but sign-up/start-page traffic plateaus—limited mainstream escape so far
- •Manus succeeded via reliability and cross-platform autonomy (email, web, slides, spreadsheets)
- •Horizontal agent startups may struggle against Big Tech distribution once capabilities commoditize
AI beyond the prompt box: desktop apps, AI browsers, and measurement challenges
They shift to new AI surfaces—especially desktop apps and AI-native browsers—and how that changes how success should be measured. Desktop-first tools can be huge revenue businesses while appearing small on web metrics, while AI browsers face high switching costs and still need a “killer feature” for mainstream adoption.
- •Desktop apps (e.g., Granola, dictation tools) are increasingly central to daily AI workflows
- •Ranking methodology challenge: web and mobile are measurable; desktop usage is harder to track
- •Future reports may need stronger revenue-based ranking to capture desktop-first winners (e.g., Cursor)
- •AI browsers: Comet led early; Atlas and others show the trend, but switching browsers is hard
- •Download spikes: Comet’s download-page interest cited ~5× Atlas, despite ChatGPT’s huge audience
How teenagers actually use AI—and why it predicts mainstream behavior
Olivia uses teen behavior as a leading indicator for consumer AI adoption, citing new survey data. Homework is now openly mainstream, while creative editing and casual/emotional conversation are rising and likely to become ubiquitous; agents will spread too, but users won’t label them as such.
- •Survey signal: teens use AI for homework (majority admit; implied much higher in reality)
- •~38% use AI for creative work (edit/generate images/video)
- •~16% for casual conversation; ~12% for emotional support/advice
- •Agents will become like ‘dot-com’—a default capability embedded in most products, not a category
Memory as a core product advantage: personalization, context boundaries, and the end of onboarding
They close on memory and personalization as the next durable moat for AI products. Olivia argues memory can feel jarring today due to context leakage between personal/professional usage, but once properly segmented, it will make products without immediate personalization feel broken—shrinking onboarding and increasing lock-in.
- •Memory is already strong in ChatGPT/Claude; Google’s ‘Personal Intelligence’ pulls from Docs/email
- •Current risk: unintentional context crossover creates discomfort or wrong-context helpfulness
- •Key infra problem: persona/context segmentation across work and personal identities
- •Long-term expectation: products should ‘know you’ immediately; onboarding should largely disappear
- •Personal experience: value increases materially after months as the system learns preferences and context
