XNews

OpenRouter usage data & 'Cinderella Glass Slipper' adoption effect

OpenRouter usage data & 'Cinderella Glass Slipper' adoption effect

OpenRouter and community analyses reveal open-source model usage patterns and adoption psychology. - Over 50% of open-source model usage is roleplay/story/interactive fiction; programming ~20%. - Programming tasks consume >20K tokens per session; other tasks ~5K tokens. - China open-source model share grew from <2% (end 2024) to ~30% mid-2025. - Inference-style model usage rose from ~0% in early 2024 to >50% by end-2025. - 'Cinderella Glass Slipper Effect': earliest model that solves a user's pain point achieves durable lock-in despite later better models. - Mid-sized models rise while tiny models die and extremely large models face cost/latency tradeoffs. - Claude's share of programming dropped below 60% in Nov 2025; OpenAI climbed from 2% to 8% in the same measure. - Asia share doubled from 13% to 31%, with growth across China, Singapore, Korea, and Japan.

Key insights

Open-source model usage skews toward roleplay and long-token programming

OpenRouter analysis shows more than half of open-source model calls are for roleplay, interactive stories, or text games, while programming comprises about 20% of usage. Programming workloads commonly feed >20,000 tokens per session because entire codebases or large file contexts are supplied; other tasks use roughly 5,000 tokens per session.

First-solution lock-in drives long-term retention ('Glass Slipper Effect')

Data and commentary describe a 'Cinderella Glass Slipper Effect' whereby users who first find a model that precisely solves a high-value pain point form foundational cohorts that rarely migrate to later, even stronger models. The effect emphasizes experience-fit and early-market timing over absolute performance.

Shift in model footprint: mid-sized models rise; inference workloads dominate

Observers report a market shift away from tiny incapable models and hyper-large costly models toward mid-sized models with balanced latency and cost. OpenRouter usage shows inference-style models occupied ~0% in early 2024 and >50% by late 2025, indicating more multi-step and tool-using interactions.

Regional shifts: rapid growth in Asia and China’s open-source share expansion

OpenRouter-derived data indicate Asia's usage share rose from 13% to 31% year-over-year, and China's share of open-source model usage grew from under 2% at end-2024 to nearly 30% by mid-2025.

Source posts