XNews

DeepSeek, OpenRouter usage and 'Cinderella' stickiness effect

DeepSeek, OpenRouter usage and 'Cinderella' stickiness effect

OpenRouter and related analyses reveal open-source model usage trends and retention dynamics. - Over 50% of open-source usage involves roleplay/story interactions; programming ~20%. - China open-source share rose from <2% (end 2024) to ~30% (mid-2025). - 'Cinderella Glass Slipper Effect': first model to solve a user's pain point captures long-term loyalty. - Mid-size models are rising while many tiny models fail and very large models remain costly. - Programming inputs consume >20K tokens; other tasks ~5K tokens. - Inference models >50% of usage by end of 2025 (from ~0% in early 2024). - Claude’s programming share dropped below 60% in Nov 2025; OpenAI share rose from 2% to 8%. - DeepSeek-V3.2 paper reports sparse attention (DSA) and expert distillation techniques to match closed-source SOTA.

Key insights

Open-source usage skewed toward roleplay and long inference workloads

OpenRouter analysis shows more than half of open-source model usage is for roleplay, interactive narratives, and similar scenarios, while programming comprises about 20% of usage. The report also noted a rapid rise in Asian usage and a shift toward inference-oriented model calls, with inference models exceeding half of total usage by late 2025.

User retention driven by first-success ‘Cinderella’ effect, not peak performance

The 'Cinderella Glass Slipper Effect' describes how users lock into the first model that perfectly solves a high-value pain point; later, even higher-performing models struggle to displace that binding. OpenRouter's call-log analysis found foundational cohorts remain loyal due to workflow and experience fit rather than raw performance.

DeepSeek-V3.2 paper claims architecture and training advances rival closed models

DeepSeek-V3.2 technical report details a sparse attention mechanism (DSA), expert-model distillation, and synthetic task environments to improve long-context and agent performance while reducing compute costs. The paper claims parity or superiority with leading closed models on select reasoning benchmarks and competitive performance in IMO/IOI-level tasks.

Source posts