User lock-in, adoption psychology and token usage trends

Usage analytics reveal strong roleplay traffic, mid‑sized model preference, and user lock‑in patterns.
Key insights
OpenRouter token analysis shows roleplay dominates and mid‑sized models rise
OpenRouter usage data indicates over 50% of open‑model traffic is roleplay/interactive scenarios, with programming making up ≈20%. Mid‑sized models (15B–70B) gained preference for cost‑performance tradeoffs. Chinese open models rose from <2% (2024) to ~30% (2025) of open model usage. Token consumption is highest for coding tasks (20K+ tokens per session) due to codebase context feeding.
Cinderella Glass Slipper Effect: first model that fits a workflow wins
Analyses of model migration show a 'Cinderella Glass Slipper' effect: users strongly stick to the first model that materially solves their high‑value workflow, creating foundational cohorts with high switching costs even when newer models offer better raw performance.
Agentic and reasoning model traffic surged relative to single‑shot generation
Relative usage shifted from single‑turn generation to multi‑step reasoning/agentic workflows: prompt lengths increased ~4× and agentic model share rose above 50% by end of 2025, reflecting more tool use and long‑running tasks.
Compute/value dynamics may shift labor value toward know‑how owners
Analyses argue that as agent compute scales, value accrues to agents with specialized knowhow and to owners of agent compute; human time value may decline unless humans supply unique knowhow that agents require for improvement.