On Reasoning, or, Why your AI Bill is About to Explode
Reasoning models are driving token usage to unsustainable levels. Learn why and how to manage the costs.
The latest generation of reasoning models like GPT-5, o1, and Claude's thinking variants promise unprecedented capabilities, but they're quietly creating a cost crisis that most AI companies haven't seen coming.
While the industry celebrates these models' ability to "think through" complex problems, Ethan Ding aptly warns that we're witnessing a "short squeeze" where reasoning capabilities are driving token consumption through the roof, making flat-rate pricing models mathematically impossible. What used to be simple prompt-response interactions are now becoming multi-step reasoning sessions that can burn through 10x or even 100x more tokens than traditional models, often without users even realizing it.
The Token Explosion Nobody Saw Coming
The numbers are pretty nuts. As Ding documents, users on Claude Code's now-discontinued unlimited tier were consuming 10 billion tokens per month, equivalent to 12,500 copies of War and Peace. Theo's analysis of GPT-5 reveals the core problem: reasoning models don't just generate longer responses, they fundamentally change how we interact with AI.
Where ChatGPT once replied to simple questions with concise answers, modern reasoning models spend minutes planning, researching, and refining their responses. Users quickly discover the "for loop effect", setting models on continuous tasks that run for hours, burning tokens like fuel while the user focuses elsewhere. What appears to be a more intelligent model is actually a more expensive toy, with usage patterns that scale exponentially rather than linearly.
The Mirage of Reasoning
Ironically, recent research suggests that much of this expensive "reasoning" may be less valuable than it appears. Zhao et al.'s experiment reveals that Chain-of-Thought reasoning may be a "brittle mirage" that vanishes when pushed beyond the training distribution. Their controlled experiments demonstrate that what looks like genuine reasoning is often just sophisticated pattern matching, models generating reasoning paths that approximate those seen during training, but failing catastrophically on novel problems. This means companies may be paying premium prices for elaborate token-heavy responses that provide marginal actual reasoning benefits over more efficient approaches.
Apple researchers have reached similar conclusions, showing that frontier Large Reasoning Models collapse once puzzle complexity passes a threshold and even shorten their "thinking" as problems grow harder, an illusion of thinking that further underscores how fragile these reasoning traces really are.
The Business Model Apocalypse
This creates an impossible prisoner's dilemma for AI companies. Usage-based pricing would solve the cost problem but kills user acquisition, nobody wants surprise bills. Flat-rate pricing attracts users but becomes unsustainable as reasoning models enable 24/7 autonomous agents that can consume unlimited tokens. As Ding notes, even Anthropic with their sophisticated auto-scaling between Opus, Sonnet, and Haiku models couldn't make unlimited pricing work.
The math has fundamentally broken: there's no flat subscription price that can support users running reasoning-intensive workloads, yet switching to metered billing often means losing customers to competitors still burning venture capital on unsustainable flat rates.
The Solution: Intelligent Usage Management
The path forward isn't to abandon reasoning models, but to implement sophisticated usage tracking and management from day one. This is why we started Parmot, providing the infrastructure to monitor token and cost consumption across providers, with data-driven analytics for smarter limits, and sustainable pricing tiers.
Companies need real-time visibility into their true AI costs, the ability to set per-user limits that prevent runaway usage, and flexible plan structures that can accommodate both casual users and power users running extended reasoning sessions.
The winners in this new landscape won't be those offering unlimited access, but those who've built sustainable business models. Try integrating Parmot today for reliable and easy subscription plans with cost-optimized usage limits.