When it comes to the cost of AI, the first question is almost always the same: how much does a token cost? It's a logical question, especially when choosing between models. But the more complex AI systems become, the less the price of a token says about the real cost of a task.

One Request — Dozens of Steps
In an agentic system, a single request can look like this:

Sometimes there are several such cycles. One agent will pick a cheap model, make 12 calls, and repeat the operation twice. Another will use a model twice as expensive but complete the task in three calls with zero retries. Which option is cheaper? Cost per token won't tell you.

AI Economics Is Above the Token Level
You need to look higher up the chain:

The word successful is fundamental here. A workflow at €0.03 with a 60% success rate versus a workflow at €0.07 with a 98% success rate — comparing only the cost of the run tells you nothing about efficiency.

Cost per Successful Outcome = Total Workflow Cost / Successfully Completed Workflows
The key question: how much does a real business outcome cost, not just the model run? Processing an incident? Reviewing a pull request? Resolving a customer request? The conversation about token price turns into a conversation about architecture: model routing, context size, tool calls, retries, memory, and the orchestration layer.

The Token Doesn't Disappear, But It's No Longer Enough
Cost per token is still a useful technical metric. But for modern AI systems, it's no longer sufficient. The cheapest token doesn't necessarily produce the cheapest result.