Token Pricing Inflates AI Usage Costs and Incentives

The Register argues that AI billing anchored to tokens has baked inflation and perverse incentives into modern LLM platforms. The opinion coins token incremental burn syndrome (TIBS) to describe progressively rising token consumption caused by design choices, retries, and feature creep, and notes that token counting is easy to implement but poor at reflecting useful work.

The piece highlights why tokens became the default metric: lexemes are straightforward to parse and count for both prompt and response. Platforms count tokens going in and tokens coming out, and often apply simple budget checks like ntokens_left to gate usage. That simplicity creates predictable billing mechanics but also incentives that drive inefficiency. Practitioners see several measurable artifacts:

• •inflated prompt engineering to coax longer outputs
• •repeated inference retries and higher tokens per interaction
• •vendor-side additions of "slop" or metadata that increase output tokens

Alternative billing approaches to consider

• •outcome or task-based pricing tied to successful completions or business KPIs
• •compute-time or GPU-second billing, aligning cost to raw compute consumed
• •session or conversation pricing, which encourages stateful, efficient interactions
• •feature-tier pricing that charges for capabilities rather than token volume

This is a product-design and platform-economics problem, not a research limitation. Paying per token is analogous to paying programmers per keystroke; it rewards verbosity and inefficiency rather than value. For ML engineers and platform owners, token-based pricing affects architecture choices: developers may add caching, batch requests, or local models to avoid expensive API calls, shifting complexity back onto teams. For vendors, tokens remain attractive because they are auditable and easy to meter, and because customers lack mature observability to map tokens to business outcomes.

Expect incremental changes: better observability (token-to-KPI mapping), hybrid pricing experiments, and vendor features that hide token costs behind higher-level primitives. The more consequential shift will be when one major provider pilots outcome- or compute-based pricing at scale; that could reset industry norms and reduce TIBS.