Paste your system prompt. We find the timestamp, the request ID or the mutable memory block that's silently re-billing your cached prefix at full input price — and put a dollar figure on it.
Paste a system prompt on the left. Nothing is uploaded — the audit runs in this tab.
| Run an audit to model the cost. |
Prefix caching is an exact-prefix match. The first changed byte kills the cache for everything after it, and no provider raises an error when that happens. These are the patterns that break it in production.
A one-off audit finds today's bug. The bug comes back the next time someone adds a
timestamp "for context freshness". prefix-audit runs in your pipeline and
fails the PR that breaks the prefix — before the invoice explains it three weeks later.
name: prompt-cache
on: [pull_request]
jobs:
audit:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- uses: actions/setup-node@v4
with: { node-version: 20 }
- run: npx prefix-audit prompts/support.md \
--provider anthropic \
--requests 8000 \
--fail-on critical \
--max-waste 500
Any critical finding, or modelled waste above your threshold, turns the check red.
Hashes the prefix and fails when it changes, so a one-character edit can't quietly cold-start every cache.
Diffs N real renders of the same prefix and points at the exact byte where they diverge.
The web tool costs nothing and always will. If you want the prefix to stay fixed across a team and a deploy schedule, that's the paid part.
Unlimited audits, all 14 rules, the cost model, auto-fix, and the open-source CLI for local use. No account, no upload, no key.
Hosted GitHub Action, baseline drift detection, PR comments naming the offending line, and Slack alerts when a deploy drops your hit rate.
Ingest cache_read_input_tokens from your provider, chart real hit rate per feature, and alert on a 10-point drop in a rolling 24h window.
Sources for the numbers on this page: Harness State of AI in FinOps 2026 (26% of AI spend wasted, 72% hit a surprise AI bill); FinOps Foundation State of FinOps 2026 (98% now manage AI spend, up from 31% two years ago); ProjectDiscovery's engineering writeup (7% → 84% cache hit rate, 59% cost reduction, 9.8B tokens served from cache). Provider cache multipliers are transcribed from public 2026 price tables and are editable above — always verify against your own invoice.