Verify before you trust
AI API Model Tester
Check whether an endpoint serves the real model it claims. Runs in your browser, your key is not sent to our servers.
Every published result is re-run on our server before it counts, and each verdict links the exact evidence that produced it. A result you believe is wrong can be disputed.Report on DiscordOpen a GitHub issue
Why this tester is different
Each check matches a concrete signal in the reply, so we don't flag real models as fake the way statistical fingerprinting does. Read the exact code.
A result only reaches the public board after our server re-runs the whole test itself, so the rankings cannot be forged.
Every verdict shows the prompt sent, the raw reply, and the matched text that triggered it. Nothing is hidden.
- Version vagueness a genuine model confirms the vendor and tier but is often unsure of its exact version. That alone is never a fail, only marked as advisory.
- Temporary errors rate limits (429), 5xx and timeouts mark a run as unverified, never suspicious. A real endpoint having a bad moment is not penalised.
- Cloud host names naming AWS Bedrock, Google Vertex or Azure as the host is accepted, since those are legitimate ways to resell the real model.
- Protocol reshaping translating between chat-completions and the native messages format is normal gateway behaviour and is not a fraud signal when the model itself is genuine.
- One flaky probe three of four probes passing is enough to be genuine, so a single off answer does not condemn an honest endpoint.
This whole tester is open source. Read exactly which probes run and why: View the detection code
Have an idea for a better probe? Suggest one here:Open a GitHub issueReport on Discord
Open source. Verify and improve it.
The test code is open. Read exactly which probes run and why, suggest new tests, or discuss the detection on our Discord.