Tim O’Brien argues that while tokens are easy to measure and appear directly on bills, they do not answer whether an AI workload delivered value. Counting tokens is useful, but treating token consumption as a proxy for value or individual productivity is misleading and can lead to costly mistakes.
Bills show consumption, not purpose
A bill can tell you how many tokens a system consumed, which models it invoked, and how much those calls cost. What it cannot tell you is whether the system accomplished anything worth paying for. That mismatch is at the heart of the current AI "hangover": visible spending without clear evidence of business benefit.
The database analogy: numbers without context
FinOps has long struggled with databases for the same reason. We can measure storage, compute, queries, replicas, and utilization, but none of those metrics explain what the database is for, why it was designed that way, or whether its cost is justified. AI is closer to that complexity than to simple metered infrastructure: a costly component may be wasteful, or it may support a critical transaction path that demands availability and low latency.
Tokens are not a universal currency
Tokens from different models are not interchangeable. You can spend hundreds of dollars a day on an AI task, but you can’t judge whether that’s expensive without understanding the goal. A coding agent that replaces hours of engineering work, a classifier sorting low-value documents, and an orchestrator coordinating dozens of downstream tasks might all generate the same daily bill, yet they are not economically comparable.
This is why O’Brien is skeptical of the current enthusiasm for "tokenomics." Token accounting is visible and often tied directly to invoices, so it belongs in reports. But naming token-based accounting does not solve the underlying blind spots.
Important operational questions token counts don't answer
You cannot simply look at a company’s token consumption and conclude, "You should use Opus" or "You should use a cheaper model." Important questions remain:
- What is the model doing (writing code, classifying documents, recognizing speech, coordinating agents, making decisions that require human review)?
- Does a cheaper model produce more retries?
- Does an expensive model reduce downstream work?
- What are the consequences when it fails?
Those questions determine the economics; treating tokens as a common currency is a mistake.
Architecture, model choice and real costs
AI architectures are increasingly layered: a company might use a mid-tier model for coordination, a frontier model for hard tasks, a local model for sensitive data, and a specialist service for speech. Some of these are billed by token, some by the minute, seat, or request, and some run on hardware the company already owns.
As local inference and platform shifts (for example, iOS 27) change the landscape, licensing costs may become more important than token consumption.
Concrete examples and trade-offs
O’Brien gives practical illustrations: was spending $500 on an hour of Fable 5 development worth it? Maybe — you need to know what was accomplished. A system that wakes up, reads email for keywords and checks the weather probably does not need 2M tokens on OpenRouter plus a Tavily subscription; a simple Python script would cost zero tokens to run.
Anyone orchestrating large agent networks with Hermes agents, Pi harnesses, or similar technologies learns fast that running the entire process on the most expensive frontier model is a recipe for financial collapse. A billion tokens run on a high-cost frontier for a simple task versus running it on Luna or GLM-5.3 can create an error that costs tens of thousands of dollars. Model selection matters, but so do reasoning level, memory management, consolidation, and choosing the right approach for the task.
The next step for AI financial management
The next step is not to abandon token accounting. Count the tokens. But connect consumption to workflow, result quality, architecture, and business outcome. Only in that context does spend make sense.
O’Brien concludes that the FinOps community lacks a substantive discussion of "value" — instead, many are still chasing simple efficiency reports that don’t address whether the work was worth doing.
Note on an event
The article mentions AI Codecon on Monday, August 31, focused on open-weight models, self-hosted stacks, and tooling. The program includes IBM Research’s Gabe Goodhart speaking about off-loading scoped tasks to local models to minimize AI costs while maximizing agent value.
