Running a standard AI benchmark task now costs about 15 cents, roughly 97 times less than a year and a half ago, according to ARK Invest on Episode 140 of its podcast, The Brainstorm. AI inference costs have dropped by up to 97 times within approximately 18 months, with historical cost reductions exceeding 90% per year during peak periods. ARK’s analysts said this cost compression is reshaping competition among OpenAI, Anthropic, Meta, and xAI, which are increasingly competing on performance-per-dollar rather than capability. ARK predicts that as cost per task approaches low single-digit cents, enterprises could collectively save hundreds of billions of dollars on knowledge work, driven by cost curves declining at 90%-plus annually. This threshold could drive the next wave of agentic AI adoption, where AI systems handle entire workflows autonomously, though the podcast noted this future hasn’t fully arrived. The bottleneck is real-world execution, not models, as translating benchmark performance into reliable action in unstructured business environments remains the critical gap. For major AI players, the cost collapse creates a strategic dilemma, with winners likely determined by distribution, integration, and pricing strategy rather than benchmark results. ARK suggests investors should watch who builds the best business model around cheap AI, as the premium shifts to proprietary data, domain-specific fine-tuning, reliable integration, and trust from consistent real-world performance.