The model census

Every Suno clip carries a field — major_model_version — naming the model generation that made it. We have stored that field since our first capture, so we can do something plain with it: over the same set of songs the platform shows people, count the model-version mix.

We hold two corpora. The first is the /api/trending Explore fossil, frozen in September 2024 (our earlier report proved it had stopped updating long before Suno removed the endpoint's public route on 2026-07-24) — our last successful capture is dated Aug 20, 2026, but the content itself stops in 2024. The second is the discovery surface Suno actually uses now: the human-curated Staff Picks shelf, which we snapshot daily from Jul 5, 2026 through Aug 20, 2026 — 47 distinct days. One is "then", one is "now": two censuses, one ruler.

Spread the fossil's 39 clips by model version and you get the mix below. It is a still photograph — not a trend, but a cross-section pinned to September 2024. This mix is fixed for good regardless: the playlist's membership has been frozen since then. Suno removed the convenience route to it, /api/trending, on 2026-07-24, but the object itself is still readable — we just read it by its own address now.

The same ruler, applied to today's Staff Picks. Below is the stacked model-version mix of the shelf across 47 days — one column per day. In the most recent snapshot (Aug 20, 2026), the largest share is v5.5 (40.9%, 9 of 22 clips). Against a fossil stuck on v3.5, the curated shelf shows a newer set of generations.

Beyond model version, each clip carries a metadata.type — how it was made: prompt-to-song (gen), Studio export, edits (several variants, which we group), concat / extend, upsample, cover, upload. Count the fossil and the latest Staff Picks day by this field and the two corpora differ in their production path. Prompt-to-song is 28.2% of the fossil and 72.7% of the latest Staff Picks day.

One value of metadata.type deserves separating out: upload — audio a human made elsewhere and uploaded to Suno. These clips are easy to miss because they carry no model badge and an empty major_model_version; nothing generated them. On 35 of our 47 captured days, Suno's curated Staff Picks shelf held at least one. So the shelf is not purely a showcase of what the models can do — part of it is a showcase of what people brought.

Suno's curation carries a second shelf, named literally "Best of v5.5". We snapshot it daily too — but a model census over it is meaningless: the shelf holds exactly one model's work by construction. In our latest capture (Aug 20, 2026) it held 23 clips, of which 100% are v5.5.

So we do not use it as a census corpus; we treat its existence as the finding: Suno stands up a dedicated shelf to merchandise its newest model. The discovery surface is shaped by editorial taste and by product marketing — the newest model is not just used, it is displayed.

One reference table, in passing. In Suno's clip metadata a model appears sometimes as a codename (model_name, e.g. chirp-fenix) and sometimes as a version (major_model_version, e.g. v5.5). Below is the mapping we observed by aligning the two in our own data. Note the framing: this is observed, Suno does not publish the correspondence, and the codenames may change without notice.

Put the two censuses side by side and the reading is simple: anyone judging "which model generation is popular on Suno now" from the public trending endpoint is handed a 2024 model mix — dominated by v3.5. The shelf the platform actually curates has long since turned over, and it stands up a dedicated display for the newest model. We draw no verdict; we set the two rulers' readings next to each other.