What happened
The open-weights model behind our engine was trained with declared Suno coverage up to v5. In late September 2026 a user brought us a Suno v6-MINI track — an English vocal pop song, days old, generated after the model's training data was frozen and present in no training set we know of.
The numbers
| Test | Result |
|---|---|
| Local engine, on the retained MP3 (already double-transcoded Opus→MP3) | P(AI) = 0.9999 — detected |
| Production API, same file through the deployed pipeline | P(AI) = 0.99990 — detected |
The transcript of this run, the file's provenance (public Suno share page) and its SHA-256 fingerprint are on the methodology page.
What this does and doesn't prove
- · Proven: the engine generalizes to at least one released v6-MINI track, and the retained file had already been double-transcoded — generalization beyond the declared v5 coverage was an open question and is now empirically answered for this sample.
- · Not proven: coverage of every v6 variant and prompt style. One track is one track; our benchmark expansion adds v6 variants per-version with quotas, and this page will be updated — including if results get worse.
Scores are uncalibrated probabilistic signals, not legal proof of authorship.
Our rolling coverage policy
New generator releases get tested against our engine as they land, and results publish either way. The current honest state: Suno v3–v5 (model-declared) + v6-MINI (verified, n=1); Udio in scope with per-track validation publishing as the benchmark expands; YuE2 not reliably detected — see our YuE2 coverage report.
Check your own Suno v6 track free: sonicert.com — unlimited checks, no signup.