Anthropic and OpenAI’s flagship AI models cost approximately 100 times more per token than their smallest counterparts, yet widespread user behavior often defaults to these expensive options for simple tasks. This inefficiency has led to significant financial waste across major technology firms. Meta employees reportedly consumed 60 trillion AI tokens in a single 30-day period earlier this year, driven by internal leaderboards like "Claudeonomics" that incentivized high usage regardless of productivity. Tech columnist Gergely Orosz estimated this specific consumption at roughly $900 million based on list prices, though actual costs may vary due to enterprise agreements.
Uber similarly exhausted its entire 2026 AI tools budget within just four months, highlighting the rapid depletion of resources when usage is not carefully managed. Bain & Company consultant Jue Wang noted that many companies default to high-end models like Anthropic's Opus for trivial tasks such as generating emails, describing it as hiring Albert Einstein to calculate change at a fast-food restaurant. Despite per-token prices dropping more than 90% since 2023 according to Silicon Data’s Token Expenditure Index, total spending on large language models has doubled since late 2025. Economists attribute this to the Jevons paradox, where increased efficiency leads to higher overall consumption.
To combat this waste, engineers are promoting "tokenmaxxing," which focuses on maximizing the value of each token rather than the volume spent. Key strategies include routing simple tasks to smaller, cheaper models, trimming context windows to avoid "context rot," and utilizing caching and batching techniques. For instance, processing 1,000 invoices can cost about $30 on a premium model but only $0.30 on a smaller one if optimized correctly. Companies like Shopify have implemented circuit breakers to cut off access during spend spikes, while others encourage starting with free or local models before escalating to paid APIs.
The disparity between model capabilities and user habits reveals a critical gap in enterprise AI governance. While technical solutions like routing and caching exist, the primary driver of waste appears to be cultural incentives that reward raw token consumption over strategic application. The case of Meta’s leaderboard illustrates how gamification can distort economic rationality, leading to hundreds of millions in unnecessary expenditure. This suggests that without strict budgetary controls and clear guidelines on model selection, organizations will continue to face unsustainable cost structures despite falling unit prices.
From an operational risk perspective, the rapid depletion of budgets, as seen with Uber, indicates that current forecasting methods for AI infrastructure are inadequate. The Jevons paradox underscores that cheaper tokens do not automatically lead to savings; instead, they often expand usage into low-value activities. Enterprises must shift from passive API consumption to active orchestration, employing guardrails such as hard spend limits and automated escalation protocols. Failure to implement these controls risks not only financial loss but also potential security vulnerabilities when sensitive data is processed through unmonitored, high-cost channels.
