THE GLASS SLIPPER EFFECT

Here's a fairy tale that explains how you get locked in to one AI model.

OpenRouter and a16z speedrun just published their State Of AI report. Among all the data is something they've called the Glass Slipper Effect.

The metaphor works like this: Users constantly test AI models against their toughest unsolved problems - trying on shoes, looking for a fit. When a new frontier model solves something that was previously impossible, those users lock in. Hard.

And they don't switch when something better comes along. Workflows have been built. Teams have been trained. Pipelines have been restructured. All around the model that cracked it first.

The report calls these users 'foundational cohorts'. And some models build large ones: GPT-4o Mini in July 2024 and Claude 4 Sonnet's in May 2025 had retention rates of 35-40% at month five - whilst later adopters churned almost immediately.

Which also explains why, when models come to the end of their lives and are no longer available, it causes foundational cohorts no end of strife. Product-market fit in AI often means workload-model fit. And if it ain't broke...

But what's today's in-the-end-at-the-end?

Well - we like to think we're optimisers, hunting for the best. In practice, we're often satisficers. Good enough - when embedded - becomes unbeatable. And that's why 'first-to-solve' often beats 'best-in-market'.

And thank goodness. Because while optimisers keep testing and remain in prototype purgatory, satisficers ship excellent production systems and improve the lives of many customers and colleagues.