McKinsey & Company says the cost of AI intelligence is falling rapidly even as enterprise bills rise because businesses are consuming far more capability through autonomous agents. GPT-4 launched in early 2023 at $60 per million output tokens, while models with roughly comparable benchmark performance can now cost well below $1 per million output tokens in some cases. Yet agentic workloads can follow different reasoning paths and produce cost differences of up to 30 times between runs. About one-third of surveyed organizations spend more than 10% of their technology and communications budgets on AI, 60% plan to increase spending within a year and roughly one in five say AI costs are beginning to constrain operating expenses. McKinsey recommends measuring agents at the task level, including execution cost, success rate and human verification time. Companies should improve visibility into spending, match models and workflows to tasks, reduce unnecessary tool calls and loops, and manage licenses, quotas and vendor terms. The firm says the objective should be directing AI spending toward the highest returns rather than cutting costs indiscriminately.