The AI Colander: Model Churn and Falling Costs

The article exposes a tension between the hype of AI model dominance and the reality of poor customer retention. AI models retain users at rates closer to mobile games (high single digits to 40% at month five) than to social networks or traditional software (80-90%). The average frontier model holds the crown for only 41 days, making any competitive advantage fleeting. This creates a fundamental instability for startups and enterprises that build on a single model.

The concrete market shift is the inclusion of cost in benchmarks. Microsoft‘s MAI-Code-1-Flash matches Claude Haiku 4.5 on SWE-Bench Verified using 60% fewer tokens. On the Intelligence Index, GPT 5.5 and Claude Opus 4.8 score near 60, but GPT 5.5 is 28% cheaper to run. xAI‘s Grok 4.5 scores 54 and runs at $0.31 per task—60% less. The price for a given level of benchmark performance is falling roughly 10x per year across frontier knowledge, reasoning, math, and software engineering tasks.

For builders, the takeaway is to embrace model churn rather than fight it. Every 41-day cycle gives buyers more negotiating leverage as cheaper or better models appear. The sensible architecture is one that abstracts away the model provider and allows rapid switching. Investing in cross-model evaluation frameworks and flexible inference pipelines is more durable than betting on any single model. This instability ultimately benefits the community by driving down costs and accelerating capability.