---
title: TokenMaxing
canonical: "https://cloudmonitor.ai/docs/using-cloudmonitor/reports/tokenmaxing/"
description: "Ways to get more value from every unit of AI spend — caching, routing, downshifting and batching, ranked by saving."
---

:::note[Not everyone has this]
**Tokenomics** is an optional module. If you can't see it in the menu, it isn't switched on for
your workspace — ask an admin to turn it on under Settings ▸ Workspace Settings ▸ Features.
:::

**Savings & Waste** for AI. Where that screen finds oversized virtual machines, this one finds
inefficient token use — ranked by what each change would save.

## The four levers

**Caching.** Repeated context re-sent on every call is charged every time. Caching it is usually the
single biggest win and needs no change to what the application does.

**Routing.** Sending every request to the most capable model is the AI equivalent of running
everything on the largest instance size. Route the simple work elsewhere.

**Downshifting.** The same idea applied to a specific workload that has outgrown the need for a
frontier model.

**Batching.** Many providers price batched or asynchronous work below real-time. Anything that
doesn't need an answer this second is a candidate.

None of these reduce what your AI does. That's what separates them from "use AI less", which is
usually the wrong answer.

## Read it as a prompt, not a promise

:::tip[These are point-in-time estimates]
The figures here are estimates of what a change would save, not measurements of your spend. **None
of the global filters narrow them** — not period, provider, cost group, virtual tags or cost basis
— and every chip is greyed out with that reason rather than pretending to apply.

If a view shows **Token telemetry not connected**, the underlying per-token metrics need a provider
usage API that isn't connected for that provider yet. See
[what isn't available yet](/docs/reference/whats-not-available-yet/).
:::

The value here is the *ranking* and the *type* of opportunity. Treat a specific figure as an
indication of scale, and confirm any real saving on
[AI Spend](/docs/using-cloudmonitor/reports/ai-spend/) after you've made the change — cost per
million tokens is the number that will actually move.

## Who this is for

Whoever can change how the application calls the model. Unlike a licence reclaim or a rightsizing,
every item here is an engineering change — so this screen is a conversation starter with a team
rather than something a FinOps practitioner actions alone.

## Related

- [AI Spend](/docs/using-cloudmonitor/reports/ai-spend/) — what it costs today, and the unit rate to
  measure improvement against.
- [AI tool adoption](/docs/using-cloudmonitor/reports/ai-adoption/) — whether you're paying for tools
  and seats nobody uses, which is the cheaper problem to fix first.