Stop Hiring Einstein to Count Your McDonald's Change 🧠🍔

Meta employees burned through 60 trillion AI tokens in 30 days earlier this year while chasing spots on an internal leaderboard nicknamed "Claudeonomics," with tech columnist Gergely Orosz estimating the bill at roughly $900 million at Anthropic's list prices, though Meta likely pays far less. Bain & Company consultant Jue Wang told the Associated Press this summer that the same inefficiency is widespread across corporate AI use: "So many companies, so many users, default to using Opus for everything, including generating emails," she said, referring to Anthropic's high-end models. Every word sent to or returned by a model is metered in tokens, which average about three-quarters of a word and determine how quickly subscribers hit usage limits or how much developers pay on the API.

Anthropic's pricing per million tokens ranges from $0.80 for the smallest Haiku tier on prompts under 100,000 tokens (longer inputs cost five times that) to $80 for its top model, a 100-fold spread that OpenAI mirrors across its own lineup, charging between $0.10 and $10 per million input tokens. Anthropic's revenue run rate climbed from about $9 billion at the end of 2025 to more than $47 billion by May, while OpenAI is engaged in a price war to keep pace. The token-spending arms race has spread beyond Meta, with Microsoft, Shopify, and Amazon running similar token leaderboards, and Nvidia CEO Jensen Huang saying on the "All-In" podcast in March that he would be "deeply alarmed" if a $500,000 engineer did not consume at least $250,000 worth of tokens. Amazon employees later admitted to inflating their usage by running the internal AI tool on irrelevant tasks, prompting a senior vice president to tell staff: "Please don't use AI just for the sake of using AI."

Engineers have since reframed the practice as "tokenmaxxing," defined as maximizing the value of each token rather than the number spent. Three mechanics drive token consumption, beginning with the agent loop, in which a model thinks, calls tools such as web search or code execution, reads the result, and iterates, resending the entire conversation on every pass. Orchestration multiplies that cost by splitting a request across worker models called subagents, each running its own loop with its own memory. Anthropic has stated that agent teams in Claude Code can consume about seven times more tokens than a standard session when teammates run in plan mode. For users looking to cut spend without coding, habits such as selecting the smallest capable model, trimming prompt context, and avoiding unnecessary agent invocations can reduce token usage sharply while preserving output quality.