Can this tool prove an API uses a downgraded model?
No. It surfaces context drift, metering anomalies, and first-turn cache hits; those signals need call-log review and repeated runs.
LLM API Audit & Integrity Inspector
Run a multi-turn audit of any OpenAI- or Claude-compatible API in your browser: context consistency, cache-hit sanity, token accounting, and hidden instruction injection. The key is sent only to the endpoint you enter.
What it checks
- Context consistency across a growing conversation
- Cache behaviour: whether cache hits agree with the usage fields
- Token accounting: visible output versus reported completion tokens, excluding reasoning tokens
- Hidden instruction injection: reported input compared with the upstream count_tokens baseline
How a run works
The page sends 3 or 8 growing turns from your browser, each run prefixed with a unique nonce. It records usage, latency, and the raw request and response for every turn, then summarises the run as an S/A/B/D grade and a score.
How to read the result
The grade is a heuristic audit. It surfaces anomaly signals but cannot prove that an API serves a downgraded model. First-turn cache hits and metering gaps should be verified against call logs; a single run is not a verdict.
Privacy and cost
The key stays in page memory and is sent only to the endpoint you enter, never to this site. A run makes several real API calls and bills at the target provider rates.
FAQ
Why is a first-turn cache hit flagged?
Every run prefixes the system prompt with a random nonce, so a normal prefix cache cannot hit on turn one. A hit there indicates cross-request caching worth reviewing.
How much does a run cost?
Three or eight short requests with a 256-token output cap, billed at the target provider rates, usually a few cents at most.