Anthropic Cuts Token Use 98.7%
Anthropic’s code-execution paradigm cuts AI-agent token use 98.7%, slashing cost and latency.
“AI Disruption” Publication 8100 Subscriptions 20% Discount Offer Link.
The current biggest bottleneck for Agents is the context window. When an Agent needs to connect to hundreds or thousands of external tools, the massive amount of tool definitions and intermediate data results quickly overwhelms the context, leading to skyrocketing costs and plummeting efficiency.
Anthropic has provided a detailed explanation of a new paradigm called “code execution,” which builds upon the Model Context Protocol (MCP) and aims to fundamentally solve the efficiency problems of AI Agents.
The core idea is simple: Stop having the model directly call tools, and instead have it write code to call tools.
Anthropic’s data shows that this shift can reduce token consumption for processing tasks from 150,000 to 2,000, saving up to 98.7% in costs and time.
So, how does this new paradigm work? And what changes does it bring?