Imagine your team uses AI to prepare a weekly report. It reads the source material, writes a draft, then revises it twice. You see one report. Behind it may be several model requests, each reading and writing.
That is the starting point for token economics. A token is a small chunk of text. Input is what the model reads, output is what it writes. With per-token billing, a task’s AI bill adds up across all those requests, including retries.
Why It Matters
A short prompt does not necessarily mean a cheap request. Anthropic’s documentation says Claude Code sends conversation history with each request. Repeated content can receive a discounted cached rate, but it still counts toward usage.
Opus 5.5’s input and output rates are twice Sonnet 5.5’s. But twice the price per token needn’t mean twice the cost per finished task, since the models may need different amounts of work.
The managerial lesson: measure what it costs to produce work your team can use. A lower AI bill can be a poor trade if employees spend longer correcting the output.
The Decision
Start with two habits: begin a fresh session when moving to unrelated work, and choose a model suited to the task. Anthropic suggests Sonnet for most coding work and Opus for complex reasoning. Check quality before changing the team’s default.
Beyond that, a further saving has to justify the time to test and maintain it.
If your team stays within an included allowance, reducing tokens alone may not lower the current bill. It becomes more valuable when it prevents limits, extra charges or a plan upgrade.
What To Do This Week
- Choose one recurring AI workflow. Record its usable outputs and roughly how much employee time goes into checking and correcting them.
- Confirm how it is billed: an included allowance, per-token charges or both. Ask its owner for last month’s usage or spend on it, estimated if reports only show it per person.
- Ask the owner for one change and its estimated monthly saving. If that saving is worth the time to test and maintain it, try it on similar tasks and keep it if the overall result improves.
What Not To Do
Don’t treat fewer tokens as proof of greater efficiency. Your team may simply be doing less work. And don’t remove instructions or context the task needs just to shrink the request.
Signal Boost
Token Economics Starts With What a Session Carries, A-Eye Level, October 19, 2026. Its routine shows how to run the test in step 3.