🔌 AI Music Generator APIs: What Actually Ships Today

If your product needs a finished song with vocals out of an HTTP endpoint today, there is exactly one major answer: ElevenLabs. Suno, the platform most people mean when they say "AI music", has no public API at all, and everything else programmatic is either instrumental-only or unofficial. That's the whole market map; the rest is detail worth getting right before you architect anything.

I test music models by listening for where they fall apart — a vibrato that wobbles on the sustain, a breath that loops. APIs I test the same way, except the artifact is usually in the terms page.

By the numbers#

• ElevenLabs Music API: $0.30 per minute of generated audio (a 3-minute song ≈ $0.90), with custom music finetunes at $1.50 each; pay-as-you-go needs no subscription, and volume tiers run free → $99/mo Pro → $330/mo Scale (ElevenLabs; our full API guide has the request anatomy)
• Suno: zero public APIs. CPO Jack Brody announced a developer API partner program on July 1, 2026 — curated intake, no self-serve access, no published pricing or timeline (our tracking guide)
• Stable Audio 3 Small/Medium: open weights, $0 marginal cost self-hosted, commercial under Stability's Community License below $1M annual revenue; Medium generates up to 6:20 tracks, and inference on an H200 runs 1.31s (Stability AI)
• Stable Audio 3 Large is not open-weight — API, fal.ai, or enterprise self-hosting only (Stability AI)
• No Stable Audio variant generates vocals. None.

ElevenLabs: the API that exists#

Per-minute billing is the quietly radical part. A credit bundle makes you pre-buy inventory and watch it expire; $0.30 a minute means a 3-minute song costs $0.90 and re-rolling its chorus twice barely registers on an invoice. For a pipeline that generates, evaluates, and regenerates (which is every pipeline that ships listenable output), that pricing shape matters more than the headline rate.

I'll flag what I listen for before trusting any generated track in a paid product: sustained notes (where synthetic vibrato gives itself away), transitions between sections, and whether a second generation of the same prompt drifts in mix balance. Run those checks on your own material through the API's cheapest path before committing to volume tiers. The craft side of prompting this model lives in our Music v2 breakdown.

The other half of the pitch is provenance: ElevenLabs states its music model trained exclusively on licensed data, cleared for commercial use and backed by deals with Merlin Network and Kobalt, while Suno and Udio still face active label litigation. For a product with revenue attached, that's not marketing copy. That's your legal exposure differential.

Suno: an announcement, not an endpoint#

The July 1, 2026 partner-program announcement was real news: the first signal that the app-only model opens up. It is also, today, just an intake form. No pricing, no timeline, no self-serve tier, and Suno's terms haven't blessed the third-party wrappers filling the gap; those reverse-engineer the app and can break, or get your account actioned, without notice. We track the program's status, and how to write a partner application that reads like a partner rather than a scraper, in the Suno developer API guide.

Building a product on an unofficial wrapper of a platform mid-litigation is a bet I wouldn't put a release date on. Prototype with one if you must; ship on one, no.

Stable Audio 3: the API you host yourself#

For instrumental work, the open-weight route deletes the per-call meter entirely. Small and Medium download from Hugging Face, run on your GPU, and generate as much as your electricity bill tolerates — up to 6:20 per track on Medium, with LoRA fine-tuning on your own reference material and inpainting to regenerate one section against picture. Commercially free under the Community License until $1M annual revenue, at which point Stability wants an enterprise conversation; Large never self-hosts below enterprise at all, living behind the API and fal.ai instead.

The absence to plan around: no vocals, in any variant. Pipelines needing beds, cues, and ambience can standardize on this and pay nothing marginal. Pipelines needing a song cannot. Setup is covered in the local install guide, and the model's broader shape in the Stable Audio hub.

Choosing without regret#

One operational habit before any of this scales: log cost per accepted track, not per generated minute. Real pipelines re-roll two or three times per keeper, so your effective rate is a multiple of the rate card, and it's the multiple you should be budgeting against. Set a spend alert on day one; per-minute billing is honest, but it is still a meter.

Match the API to the deliverable, and be honest about the deliverable:

• Songs, with vocals, from an endpoint, this sprint: ElevenLabs. Only door that's open.
• Instrumentals at volume with infra control: self-hosted Stable Audio 3 — $0.90/song beats $0 only until your volume says otherwise.
• Suno specifically: apply to the program, build the fallback first.
• The "free AI music API" search that led here: ElevenLabs' free API tier for evaluation, Stable Audio self-hosted for production — and re-read the $1M clause before your product succeeds its way past it.

Everything above traces to the vendors' own pages and announcements as of August 2026, and this market is moving — the full generator landscape is the wider map. When Suno's program produces a real, priced, public endpoint, the calculus here changes overnight. Until then, architecture decisions built on announcements are just hope with a repo.