--- title: "The AI Colander" description: "AI is like a colander. Customers pour in, \u0026 many pour out. Retention lands between a mobile game \u0026 a social network, while the price of intelligence keeps falling." categories: ["AI"] keywords: ["AI retention","model switching","intelligence per dollar","OpenRouter","Artificial Analysis","frontier models"] ai_summary: "AI customer retention sits between mobile games and social networks. Frontier models keep the crown for about 41 days, and the price of a given level of intelligence falls roughly 10x per year, giving buyers more leverage every cycle." date: 2026-07-13 lastmod: 2026-07-31 canonical_url: https://www.tomtunguz.com/the-ai-colander/ author: "Tomasz Tunguz" --- AI is like a colander. Customers pour in. Many pour out. How many customers does a model retain after one month, two months, or three months? The answer lands somewhere between a social network & a mobile game. {{< email_image src="lrx6dfzt5aivmmfh1jru" alt="What do you have with half an AI? Cohort retention across AI models, SaaS, social media, & mobile games" width="540" height="360" >}} Software companies retain roughly 90% of customers through the first five months, an aggressive assumption. Facebook & Instagram hold closer to 80%. A typical mobile game keeps a few percent. Models land somewhere between high single digits & about 40%, with the stickiest foundational cohorts near the top of that range at month five.[^openrouter] The average frontier model maintains its crown for 41 days.[^darwin] The king is dead, long live the king, is a nearly monthly chant. This pattern isn't changing. Two days ago, OpenAI published this chart showing a surge of users for its newest model, Sol. The token growth chart is even steeper. {{< email_image src="dl2jjrnenbyjxvnphzl6" alt="OpenAI usage chart for Sol, July 2026" width="540" height="666" >}} Benchmarks are no longer purely performance-based. They include cost. Microsoft's MAI-Code-1-Flash matches Claude Haiku 4.5 on SWE-Bench Verified using 60% fewer tokens.[^mai] For the financial readers, this is the equivalent of the PEG ratio[^peg], where value is normalized by another metric. Here, intelligence is normalized by dollar. {{< email_image src="c5rrwkbzaljy0brquxzf" alt="Artificial Analysis : Cost per Intelligence Index Task, 12 Jul 2026" width="540" height="259" >}} Artificial Analysis measures the same phenomenon at the frontier. GPT 5.5 & Claude Opus 4.8 score within a point of each other on the Intelligence Index near 60. GPT 5.5 is 28% less expensive to run.[^aa-index] xAI's Grok 4.5 lands one tier below, at a score of 54, & runs the Intelligence Index at $0.31 per task, 60% less.[^aa-grok] The price for a given level of benchmark performance is falling roughly 10x per year on frontier knowledge, reasoning, math, & software engineering tasks.[^aa-price] Unlike the instability of the Middle Ages, this war of the weights benefits the community. Startups build on whichever model wins this month & re-route next month. Customers gain negotiating leverage every 41 days. [^openrouter]: OpenRouter & a16z, *State of AI : An Empirical 100 Trillion Token Study with OpenRouter* — [^darwin]: Tomasz Tunguz, *Darwin's Finches in AI* — [^mai]: Tomasz Tunguz, *Intelligence Per Dollar* — [^peg]: PEG ratio (price/earnings-to-growth) : a valuation metric that divides a stock's price-to-earnings multiple by its earnings growth rate, normalizing valuation by growth. [^aa-index]: Artificial Analysis, *Intelligence Index* — [^aa-grok]: Artificial Analysis, *Grok 4.5 brings xAI to the intelligence frontier* — [^aa-price]: Artificial Analysis & Epoch AI, *The Price of Progress : Price Performance & the Future of AI* —