# Calculate prompt caching savings without double-counting tokens

Cached tokens are not generally free. A useful caching estimate separates the tokens that were actually read from cache from fresh input, cache writes and output. The provider's usage response is the starting point; repeated prompt text alone does not prove a cache hit.

## Calculate one request

For rates quoted per million tokens:

```text
cost = (fresh_input × input_rate
      + cache_reads × cache_read_rate
      + cache_writes × cache_write_rate
      + output × output_rate) / 1,000,000
```

This is a simplified calculation. Service tiers, long-context thresholds, residency and different cache-write lifetimes may change the applicable rates. Use the [model's rate table](/models) and [price-book rules](/docs/price-book), then compare to the provider bill.

Consider an illustrative workload with 2,000 fresh input tokens, 8,000 cache-read tokens and 500 output tokens. Suppose the rates are $3 input, $0.30 cache read and $15 output per million tokens:

| Part | Calculation | Cost |
|---|---|---|
| Fresh input | 2,000 × $3 / 1M | $0.0060 |
| Cache read | 8,000 × $0.30 / 1M | $0.0024 |
| Output | 500 × $15 / 1M | $0.0075 |
| Total | Sum of the three parts | $0.0159 |

At the same hypothetical rates, treating all 10,000 input tokens as fresh would cost $0.0375 including output. The read saves $0.0216 on this request, about 58% of the total. That is not the same as a 90% discount on the entire request. These numbers illustrate the method, not a current model quote.

## Avoid the two common mistakes

First, check whether a provider's reported input total already includes cache reads. OpenAI-shaped prompt totals include cached input, so subtract the cached count before passing fresh input separately. Veqen's wrappers handle supported provider usage formats for you.

Second, count cache creation. A reusable prefix may have a write charge before later reads save money. If the extra write cost is W and each reuse saves S, the cache needs more than W / S reuses to produce net savings. Failed hits, expiry and changing prefixes can reduce that benefit.

```python
import veqen

# Values from the provider's usage response, not tokenized prompt strings.
cost = veqen.cost(
    "gpt-4o",
    input_tokens=2000,
    output_tokens=500,
    cache_read_tokens=8000,
)
```

This uses the locally installed SDK's price book. The hypothetical rates above are separate from this real model lookup. See [Python pricing](/docs/python) for supported cache-write and service-tier arguments.

## Measure by feature and customer

A site-wide cache-hit average hides the product changes that matter. Split cost and cache usage by `feature`, then by customer or project. A common system prompt may cache well while customer-specific document retrieval changes on every request. Preserve useful context; do not sacrifice answer quality just to improve the hit rate.

Veqen's [Savings and Insights](/docs/insights) can identify caching opportunities from recorded usage. Estimates describe an opportunity, not a provider guarantee. Use the [simulator](/simulate) for a workload with explicit fresh and cached inputs, and compare actual cost before and after a controlled change.

## Check the financial result

Use the same date range and billing host when comparing the request ledger to the [provider's bill](/docs/reconciliation). Credits, negotiated discounts, untracked traffic and delayed reports can explain differences. A cache estimate is not proof that an invoice decreased by that amount.

Next: [cost per customer](/guides/cost-per-customer), [agent cost per run](/guides/ai-agent-cost-per-run), and [model pricing](/models).