# Tokens Aren't Dollars
The AI industry has developed a dangerous habit. It measures progress, cost, and complexity through token counts, treating each token as a unit of economic value. Tim O'Brien's analysis, originally published on his Medium blog, cuts through this misconception and exposes why equating tokens to dollars obscures the actual economics of large language models.
A token represents a fragment of text. It could be a word, part of a word, or a punctuation mark. Different models tokenize differently. OpenAI's GPT models, Anthropic's Claude, and Meta's Llama each break down text at different granularities. What costs one token in one model might cost three in another. Charging by tokens makes billing transparent and measurable, but it creates a false equivalence between token volume and computational effort.
The problem runs deeper than accounting mechanics. Token counts hide what actually happens inside a model. A complex reasoning task might consume fewer tokens than a repetitive, simple query. A task that touches rare vocabulary or requires deep context synthesis can demand vastly more processing power per token than straightforward text completion. Yet both transactions appear identical on a bill measured in token volume.
Consider a customer running two workloads. The first generates marketing copy, feeding 10,000 tokens into Claude and receiving 2,000 tokens back. The second runs a financial analysis across historical documents, consuming 8,000 tokens of input and producing 1,500 tokens of output. The token math suggests similar computational intensity. The reality differs sharply. The financial analysis triggers multiple forward passes through the model's attention layers, complex numerical reasoning, and careful context weighting. The marketing copy flows through inference more efficiently. Yet both cost roughly the same amount if token pricing is flat.
Infrastructure costs don't scale linearly with tokens either. A model's operational expense includes GPU time, memory bandwidth, storage, and engineering overhead. These costs fluctuate based on model size, inference speed, cache efficiency, and data center utilization. Tokens don't capture batch size effects, where processing multiple queries simultaneously reduces per-token overhead. Tokens don't account for cache hits, where recent context improves throughput without proportional computational cost.
This matters because token-based pricing creates perverse incentives. Companies optimize for token minimization rather than user value. A system that could solve a problem more elegantly by asking a model to think through multiple approaches gets punished for higher token consumption. Verbose, step-by-step reasoning becomes expensive compared to risky shortcuts. Users learn to squeeze their prompts rather than articulate their needs clearly.
The cloud computing industry solved similar problems decades ago. AWS moved from simple compute-hour billing to more nuanced models that account for memory, bandwidth, and storage separately. Modern infrastructure pricing reflects the actual bottleneck costs in running distributed systems. AI pricing remains primitive by comparison.
O'Brien's central argument holds: tokens serve as a convenient billing mechanism, but they're not a valid proxy for value delivery, effort, or cost. The industry should develop more sophisticated models that reflect actual computational requirements. Until it does, both providers and customers remain trapped in a system that measures activity rather than outcome.