> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tensormesh.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Cache Savings

> See how much Tensormesh's KV caching is saving you across your serverless inference.

Navigate to **Management → Cache Savings** to see a dollar breakdown of what Tensormesh's KV caching is saving you over time.

***

## Understanding Your Savings

Cache savings = money you didn't spend because tokens were served from the KV cache instead of being recomputed.

* Every request that reuses a cached prefix (system prompt, shared document, conversation history) costs **\$0.00** for those tokens
* The savings figure shows what those cached tokens would have cost at the standard input rate
* A growing savings number means your prompts are well-structured and your cache is being used effectively
* Low savings relative to input spend usually means requests aren't sharing consistent prefixes

***

## How Savings Are Calculated

**Estimated cache savings = cached token count × standard input rate for that model**

Example: 1,000 requests each reusing a 2,000-token system prompt → 2,000,000 tokens saved. The higher your request volume and the more consistent your prompts, the faster savings compound.

***

## Maximizing Your Savings

<AccordionGroup>
  <Accordion title="Use consistent system messages" icon="bullseye">
    Keep your system prompt identical across requests. Even a single character change creates a cache miss.
  </Accordion>

  <Accordion title="Put stable content first" icon="layer-group">
    Structure prompts so that static content (system prompt, shared context) comes before variable content (the user's latest message). The cache matches from the start of the prompt.
  </Accordion>

  <Accordion title="Subscribe to External Storage" icon="database">
    [External Storage](/external-storage) persists your KV cache across sessions — dramatically increasing the fraction of requests that hit the cache for returning workloads.
  </Accordion>

  <Accordion title="Monitor hit rate by model" icon="chart-line">
    Check for per-model cache hit rates. Low hit rates on specific models often signal inconsistent prefix structure in those request flows.
  </Accordion>
</AccordionGroup>

***

## Related

* [Pricing Overview](/pricing-overview)
* [Serverless Usage](/serverless-usage)
* [External Storage](/external-storage)