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

# Trace caching

When you evaluate the **same trace under multiple criteria**, Composo reuses the
shared part of the prompt instead of re-processing it every time — and bills
those reused tokens at a **reduced rate**.

## How it works

Most evaluation workloads send **one trace, several criteria**: the same agent
conversation is scored against criterion A, then B, then C — each as a separate
request. The criteria differ, but the trace is identical across all of them.

Composo caches that shared trace:

* The **first** evaluation of a trace primes the cache.
* Any **subsequent** evaluation of the **same trace** within **30 minutes** reuses
  it. In the one-trace-many-criteria pattern, that means **every criterion after
  the first is cached**.

The cache key is the trace itself. The per-request criterion is *not* part of it
(that's the thing that changes each call), so varying the criterion doesn't break
the cache — which is exactly why "one trace, N criteria" benefits.

## What it costs

Cached trace tokens bill at **40% of the normal per-token rate** — a **60%
discount** on the reused portion.

Concretely, a request's billable tokens are:

```
billable_tokens = tokens_used − 0.6 × cached_trace_tokens
```

So for a trace evaluated under 4 criteria within 30 minutes, the trace tokens of
the 2nd, 3rd, and 4th calls are each discounted by 60%. Only the trace is
discounted — the criterion and other per-request content always bill at the full
rate.

<Note>
  Caching never changes your scores, your responses, or which requests succeed. It
  only affects how the reused trace tokens are priced. The first call on a trace is
  always billed in full (there's nothing to reuse yet).
</Note>

## Getting the most out of it

* **Batch criteria for the same trace close together.** Send the criteria for a
  given trace within the 30-minute window so the 2nd onward land on the cache.
* **Reuse traces verbatim.** The cached trace must be byte-for-byte identical;
  any change to the messages starts a fresh trace.
* It applies to the `align-20260109` model core (the current default). Lightning
  variants don't participate.

## Seeing your caching

The `/usage` page has a **Trace Caching** card and a **Cached tokens** column in
the per-model table:

* **Tokens reused** — the share of your cacheable trace tokens that were served
  from cache.
* **Cache hit rate** — the share of your evaluations that reused a cached trace.
* **Cached tokens** (per-model table) — the trace tokens that earned the discount.

The credits shown already reflect the discount, so your **Credits** column is what
you actually pay after caching.

## Questions

Reach out to [support@composo.ai](mailto:support@composo.ai) — we're happy to walk
through your caching numbers or how to structure requests to benefit from it.