Morgan Stanley points out that while AI model prices continue to decline, agent applications are significantly increasing computing power consumption. The price of mainstream model output tokens has dropped from approximately $60 per million to $10-20, a decrease of up to 83%. However, agents need to repeatedly reason, call tools, execute code, and verify results, with the token consumption for a single task being approximately 5-30 times that of ordinary chat. The bank estimates that if a complex enterprise task is completed collaboratively by 5-7 agents, each agent consumes approximately 15 times the tokens used for ordinary queries, resulting in a total task consumption of 375,000-525,000 tokens. OpenAI data shows that as of June 2026, Codex output tokens accounted for 64% of the total output tokens of its enterprise clients' Codex and ChatGPT products. This means that the decline in token price may be offset by the increase in call volume and task complexity; the price reduction of models may not necessarily reduce overall computing power requirements, but may instead further increase computing and power consumption by expanding the scale of applications.
