The LLM Productivity Cliff: Threshold Productivity and AI-Native Inequality

We define the LLM productivity cliff as a threshold phenomenon: users who adopt an engineering mindset for working with LLMs attain step-change productivity, while others experience modest gains or outright slowdowns. Synthesizing 2025 evidence across software development, customer support, and labor markets, we show that outcomes are diverse and discontinuous, with performance grouped above and below a capabilities threshold. We operationalize this threshold as architectural literacy, not as marginal prompt skill but as a qualitative shift toward decomposition, orchestration, and systematic validation. We identify boundary conditions that make cliffs more likely (task complexity, scaffolding, mental models) and develop a three-level account of inequality at the individual, organizational, and market levels. We conclude with interventions to reduce capability disparities: embedding scaffolding in tool design, institutionalizing architectural literacy training, and promoting equitable diffusion of architectural literacy and AI-native organizational practices.