Originally published on Medium, 18 July 2026. Republished here with light editing.
Suno’s public trending endpoint has been serving a September 2024 snapshot for about 22 months. I watched it not move for 84 straight days.
There is an endpoint on Suno’s public API called /api/trending.
The name implies something alive — a feed that churns, a chart that turns over, the sound of what people are making right now on the largest AI music platform in the world. A company that just raised $400 million at a $5.4 billion valuation, with roughly two million paying subscribers, generating an amount of music per day that would have been the entire output of the recording industry a generation ago.
The endpoint returns 49 songs. Every single one of them was created between September 15 and September 17, 2024.
It has not changed since.
What I was actually doing#
I did not set out to find this. I was building something small and slightly silly: a service that renders your Suno songs as animated SVG cards for your GitHub profile README. To do that I needed to read Suno’s public data — the same anonymous, unauthenticated GET requests your browser makes when it loads a song page.
While I was in there, I started snapshotting what I saw. Every day, automatically, committed to a repository. Not because I had a thesis, but because instrumenting things is cheap and past-you is always stingy with data.
I started on April 12, 2026. By July 5 I had 84 consecutive daily captures of the trending endpoint.
The top ten had turned over zero percent.
Not “slowly.” Not “less than I expected.” Zero. Same songs, same order, every day, for three months. The number one clip gained 56 plays across the entire window — roughly one play every other day, which is approximately what you would expect from a song that nobody is being shown and that a handful of people stumble into by accident.
That is when I stopped treating it as a data source and started treating it as a question.
Three hypotheses, and how to kill two of them#
A frozen feed has three plausible explanations, and they are not equally interesting. One: an editorial freeze — Suno curated this list, liked it, and stopped updating it. Two: a stale cache — some CDN edge node pinned an old response and never revalidated. Three: an abandoned object — the endpoint points at a data record that the company stopped writing to and forgot about.
These sound similar. They are not. The first means the list is intentional. The second means the list is a plumbing bug. The third means the list is a fossil — and that anyone consuming this endpoint, including every third-party tool and every researcher who assumed “trending” meant trending, has been reading a document from the deep past and calling it the present.
On July 5, 2026, I ran the forensics.
Test one: what do the objects say about themselves?#
Every clip in the response carries a created_at field. All 49 clips fall in a 72-hour window: September 15 to 17, 2024. Not “mostly.” Not “with a few older ones.” All of them, in a three-day band, twenty-two months before the request.
A live trending chart on a platform that generates millions of songs a month does not, by chance, contain only songs from one specific weekend in 2024.
Test two: is something between me and the origin lying?#
I fetched the endpoint twice, 108 seconds apart, and compared the responses. Byte-identical. That alone proves nothing — a cache would do exactly that.
So I read the headers. Cf-Cache-Status: DYNAMIC.
That is the tell. DYNAMIC means Cloudflare did not serve this from an edge cache. It means the request went to the origin, the origin ran, and the origin returned this. The freeze is not in the network. The freeze is in the database. There is no stale cache to bust, because nothing is stale — the origin is faithfully serving a record that simply has not been written to since 2024.
Hypothesis two is dead.
Test three: does the live product agree?#
On May 14, 2026, Suno shipped an “Explore Refresh” — the current, live discovery surface, built on curated playlists. Staff Picks. Best of v5.5. Real shelves with real recent content; when I checked them, they carried songs from June 2026, days old.
I compared the membership of those live shelves against the 49 clips in /api/trending.
Zero overlap. Not one song in common.
Whatever /api/trending is, Suno’s actual, shipping, user-facing Explore has no relationship to it. It is not a stale view of the live thing. It is a different thing entirely — one nobody is looking at.
Hypothesis one is dead. This was never an editorial choice. Nobody chose these songs in 2026. Somebody chose them in September 2024, and then everyone went home.
The correction I had to make to my own work
Here is the part I would prefer to leave out.
Before July 5, I had written up an earlier interpretation of this data. I had it wrong. I had concluded the list was “origin-pinned but live” — that Suno was serving a deliberately frozen chart, an editorial decision. I built an entire instrument to measure the cadence of a feed I had already misclassified. I wrote analysis on top of that instrument. I had a verdict field in a generated report that read: INDETERMINATE — expect a confident call within 1–2 weeks of daily runs.
No more days were going to help. I was waiting for motion from a corpse.
The forensics on July 5 corrected the interpretation and retired the instrument in the same commit. I mention this because the entire value of publishing an observation is that someone can check it, and you cannot ask for that credibility while quietly deleting the version you got wrong.
There is also a genuine red herring worth naming: there is a second thing on Suno’s surface called “Trending” — a curated shelf, distinct from /api/trending. It is also stale, frozen around May 2025. I spent real time confused by this. If you go looking, you will find two different frozen “trending” objects from two different eras, which is either a coincidence or a small essay about how discovery features get built and abandoned at a fast-moving company.
Why this is worth 1,500 words#
Because “trending” is a claim about reality, and this one has been false for almost two years.
Three things follow, in ascending order of how much I think they matter.
One: anything downstream of this endpoint is reporting fiction. If any tool, dashboard, tracker, or analysis has been reading /api/trending and presenting it as current Suno activity — and the endpoint’s name makes that a completely reasonable thing to have done — it has been showing its users a snapshot from September 2024. I know at least one system that did, because it was mine.
Two: this is what platform archaeology looks like. Companies moving at Suno’s speed leave objects behind. A feature ships, gets replaced, and the old endpoint keeps answering — no error, no deprecation notice, no 410 Gone, just a polite 200 OK and a cheerful payload of ancient data. The endpoint is not lying. It is doing exactly what it was built to do. It is just that the thing it was built to do stopped being meaningful in 2024, and nothing in the system is capable of noticing that.
Three, and this is the one I actually care about: the AI music era has almost no historical record, and it is being generated at a rate that makes the gap grow every day. We have detailed archives of what topped the charts in 1964. We have essentially nothing about what was popular on the largest AI music platform in the world in September 2024 — except, apparently, this: 49 songs, frozen mid-stride, still being served by an endpoint that forgot to stop.
That is an accidental time capsule. Suno’s engineers did not mean to build one. But if you want to know what the machine thought was good in the week of September 15, 2024, there is exactly one place left that will tell you, and it is a URL that nobody at the company has thought about in twenty-two months.
I find that genuinely moving, and I would rather write it down than let it get quietly patched.
A note on last week’s breach#
I am publishing this four days after 404 Media reported that a hacker breached Suno and exposed internal source code detailing what the company scraped to train its models. I want to be unambiguous about the relationship between that story and this one: there isn’t any.
I have no breached material. I have not seen any of it. Nothing in this post comes from anything other than public URLs that anyone can request from an ordinary browser, and every claim here was reproducible months before that breach happened — the capture log starts in April.
It would be convenient to ride that news cycle. I would rather draw the opposite line, because I think it is the more interesting one: you did not need to hack Suno to learn something true about it. The largest AI music platform in the world has been serving a two-year-old snapshot from an endpoint named “trending,” in the open, to anyone who asked, the entire time. Nobody looked. That is a finding about attention, not about access.
If anything, the breach makes the boring method matter more. Anyone can check my work without touching a single thing they are not entitled to see.
Method and reproducibility#
Everything above comes from anonymous, unauthenticated GET requests to public endpoints — the same requests an ordinary browser makes when it loads a public Suno page. No login, no credentials, no automated sign-in, no access to anything gated. I did not attempt to access private data, and nothing here required it.
The raw evidence — both fetches, the full response headers, the live-shelf comparisons, and the daily capture log — is archived with timestamps. The finding is reproducible by anyone who wants to fetch the endpoint and read the created_at fields.
If you work at Suno and would like this endpoint retired, deprecated, or pointed at something real: please do. I will happily update this post to say so. An honest 410 Gone is better than a cheerful 200 OK full of 2024.
If you cover AI music and want the underlying dataset, the capture timeline, or the methodology, it is yours — just ask.
I run Seismophone (https://seismophone.chanmeng.org), an independent observatory for AI music. It publishes what the platforms don’t. Numbers over adjectives.
— Chan Meng, Aotearoa New Zealand