Source: Azeem Azhar, Exponential View
Anthropic's release of Claude Thinking (formerly Mythos Preview) exposes a pricing arbitrage: extended reasoning—where models work through problems step-by-step before answering—produces meaningfully better outputs on complex tasks, yet most pricing models treat all inference equally. Enterprises running technical or analytical workloads can now access qualitatively superior problem-solving within existing API budgets, forcing competitors to either restructure their pricing tiers or lose differentiation. The question is whether OpenAI, Google, and others will absorb the margin hit or charge explicitly for thinking time, reshaping how organizations budget AI labor replacement.