Model guide
The AI music models, and what you're allowed to do with them.
Who makes each model, how you can get at it, what it costs, how long and how editable its output is, and the licence that decides whether you can use it. Every row links to its source; checked October 28, 2026.
| Model | Maker | Access | Vocals | Max length | Price |
|---|---|---|---|---|---|
| Developer APIs | |||||
| Eleven Music (v1, v2, v2.5) | ElevenLabs | Self-serve API on paid plans | Yes | 5 min in the capability docs; the API accepts up to 10 min | $0.15 per minute of output (list) |
| Stable Audio 2.5 | Stability AI | Self-serve API; also on fal, Replicate and on-premises | No | 3 min | 8 Stability credits ($0.08) per generation |
| Stable Audio 3.0 | Stability AI | Self-serve API (asynchronous: submit, then poll; results kept 24 h) | Not verified | Over 6 min (Large); 6:20 (Medium) | 8 Stability credits ($0.08) per generation |
| Lyria 3.5 | Google DeepMind | Gemini API, paid tier only | Yes | About 3 min ("a couple of minutes" in the docs) | $0.08 per song |
| Lyria RealTime | Google DeepMind | Gemini API, experimental | No | Continuous stream | Not published |
| Song-generation apps | |||||
| Suno (v6) | Suno | Consumer app; no official developer API | Yes | Not published for v6 | Subscription |
| Udio | Udio | Consumer app; audio, video and stem downloads disabled since 30 Oct 2025 | Yes | — | Subscription |
| Mureka | Kunlun Tech | Consumer app and official developer platform | Yes | Not published | Per song, through prepaid tiers |
| MiniMax Music | MiniMax | Paid music API closed to new users from 20 Aug 2026 | Yes | — | — |
| Open-weight models | |||||
| ACE-Step 1.5 | ACE Studio and StepFun | Download; hosted on fal and others | Yes | 10 s to 10 min | Self-host, or about $0.0002 per second of audio on fal |
| YuE v1 and YuE2 | HKUST and M-A-P | Download | Yes | Several minutes | Self-host |
| HeartMuLa-oss-3B | HeartMuLa | Download | Yes | 4 min default | Self-host |
| DiffRhythm | NWPU ASLP Lab | Download | Yes | 1:35 or 4:45 depending on checkpoint | Self-host |
| MusicGen / AudioCraft | Meta | Download; hosted on Replicate | No | About 30 s | Self-host, or about $0.05 per run on Replicate |
| Stable Audio Open 1.0 / Small | Stability AI | Download; hosted on fal | No | 47 s (1.0); 11 s (Small) | Self-host |
| Magenta RealTime | Download | No | Continuous stream (2 s chunks) | Self-host | |
| Royalty-free music APIs | |||||
| Beatoven.ai (Maestro) | Beatoven.ai | Self-serve API keys; also on fal | No | Not published | Not published |
| Loudly | Loudly | Sales / custom pricing | No | Set by request | Custom |
| Mubert | Mubert | Self-serve | No | 25 min | From $49 trial; $199–$499/month tiers |
| Soundraw | Soundraw | Self-serve | No | 5 min | $29.99/month for 100 songs; $300/month for 1,000 |
| CassetteAI | CassetteAI | On fal | No | Several minutes | $0.02 per output minute |
Developer APIs
Developer APIs
Models a product can call directly from the company that makes them, with published self-serve pricing.
Eleven Music (v1, v2, v2.5)
ElevenLabs
Song model with vocals and lyrics. v2.5 (Sep 2026) adds composition plans with per-section prompts and lyrics, section inpainting and an audio reference.
- Editing and control
- Composition plans (sections of 3–120 s), section inpainting, audio reference, finetunes
- Licence
- Closed. Music Terms set media rights per plan; below top-tier Enterprise they exclude film, TV, radio and studio games.
Licensed from Merlin, Kobalt and Believe at launch; signed a UMG deal on 10 Sep 2026 for a separate fan platform.
Stable Audio 2.5
Stability AI
Instrumental, loop and sound-design model released Sep 2025, built for production audio.
- Editing and control
- Text-to-audio, audio-to-audio, inpainting and continuation
- Licence
- Closed. Stability says it was trained only on licensed data.
Stability has strategic alliances with UMG (Oct 2025) and WMG (Nov 2025); all three majors joined its $76M funding round.
- stability.ai/news/stability-ai-introduces-stable-audio-25-the-first-audio-model-built-for-enterprise-sound-production-at-scale
- docs.stability.ai/pricing
- stability.ai/news/universal-music-group-and-stability-ai-announce-strategic-alliance
- musicbusinessworldwide.com/universal-sony-warner-join-76m-funding-round-in-stability-ai
Stable Audio 3.0
Stability AI
May 2026 family of four models. Large is API and enterprise only; Small SFX, Small and Medium ship as open weights.
- Editing and control
- Multi-segment inpainting, continuation, LoRA training on Small and Medium
- Licence
- Large closed; Small and Medium under the Stability Community License (free below $1M annual revenue).
Lyria 3.5
Google DeepMind
Full songs with vocals and timed lyrics, in the Gemini API since Sep 2026. Single-turn: no editing after generation.
- Editing and control
- Text and up to 10 reference images; your own lyrics with section tags. No audio input.
- Licence
- Closed. Every output carries a SynthID watermark.
The Lyria 3 models it replaced were labelled legacy about seven months after launch. Google has not disclosed Lyria 3.5's training data.
Lyria RealTime
Google DeepMind
Experimental live music stream over a WebSocket, steered while it plays.
- Editing and control
- Live BPM (60–200), scale, density, brightness, drum and bass mutes
- Licence
- Closed; output is watermarked.
Song-generation apps
Song-generation apps
Consumer products that make full songs with vocals. Most have no official developer API.
Suno (v6)
Suno
The best-known song generator. Moved entirely to v6, v6-wild and v6-mini on 9 Sep 2026; Suno Studio adds a browser workstation.
- Editing and control
- Plain-language section edits, single-lyric changes, mashups, instrument isolation
- Licence
- Closed. Commercial use depends on an active paid plan.
Warner settled and licensed Suno in Nov 2025. UMG and Sony filed a second suit on 18 Sep 2026 over 60,202 recordings, including the v6 models. A Munich court found for GEMA against Suno on 31 Jul 2026 (not final).
- suno.com/blog/introducing-v6
- musicbusinessworldwide.com/warner-music-group-settles-with-suno-strikes-first-of-its-kind-deal-with-ai-song-generator
- musicbusinessworldwide.com/universal-and-sony-sue-suno-for-a-second-time-claiming-platforms-v6-models-are-the-fruit-of-the-same-poisoned-tree
- twobirds.com/en/insights/2026/germany/munich-district-court-rules-on-ai-generated-music-gema-v-suno
Udio
Udio
Song generator that settled with UMG and WMG and is building a licensed platform.
- Editing and control
- —
- Licence
- Closed.
Settled with UMG (Oct 2025) and WMG (Nov 2025). Sony filed a second suit in Jul 2026.
Mureka
Kunlun Tech
Song generator with vocals, instrumentals, stems and sheet music — the one app in this group that still sells an official developer API.
- Editing and control
- Lyrics or description input, stems
- Licence
- Closed.
MiniMax Music
MiniMax
Song models 2.0 to 3.0, now marked discontinued on MiniMax's API. Music 3 has been published as open weights.
- Editing and control
- —
- Licence
- Closed API; Music 3 weights on Hugging Face (licence not verified).
Open-weight models
Open-weight models
Weights you can download and run yourself. The licence decides whether you may use them in a paid product — and it can change between versions.
ACE-Step 1.5
ACE Studio and StepFun
Full songs with lyrics in 50+ languages from a language-model planner plus a diffusion renderer. Released Jan 2026; XL variant Apr 2026.
- Editing and control
- Repaint a region, cover, stem extraction, multi-track layering, accompaniment for a vocal, LoRA from ~8 songs
- Licence
- MIT (code and weights); the model card permits commercial use of outputs.
Training data is described (licensed, public-domain and royalty-free music plus synthetic audio) but not quantified.
YuE v1 and YuE2
HKUST and M-A-P
Lyrics-to-song models. YuE v1 (Jan 2025) is Apache-2.0; YuE2-3B (Sep 2026) outputs 48 kHz stereo under a stricter licence.
- Editing and control
- YuE2: cover via score transcription
- Licence
- v1: Apache-2.0 (attribution requested). YuE2: CC BY-NC 4.0 with a creator carve-out — companies need a licence from the authors.
HeartMuLa-oss-3B
HeartMuLa
Lyrics-plus-tags song model released Jan 2026; generates at roughly real-time speed.
- Editing and control
- Tags and lyrics
- Licence
- Apache-2.0.
DiffRhythm
NWPU ASLP Lab
Diffusion song model from timestamped (LRC) lyrics; v1.2 adds editing and continuation.
- Editing and control
- Edit, continuation
- Licence
- Apache-2.0.
MusicGen / AudioCraft
Meta
Influential 2023 research family (MusicGen, MusicGen-Melody, JASCO, MusicGen-Style). Instrumental, short clips.
- Editing and control
- Melody conditioning, style conditioning
- Licence
- Code MIT; weights CC-BY-NC 4.0 — not for commercial use, including through a hosted endpoint.
Stable Audio Open 1.0 / Small
Stability AI
Short-form sound and loop models with the best-documented training data here: 486,492 Creative Commons recordings from Freesound and the Free Music Archive.
- Editing and control
- Text only
- Licence
- Stability Community License: free below $1M total annual revenue, enterprise licence above.
Magenta RealTime
Live, streaming instrumental model trained on about 190k hours of stock music. MRT2 (230M and 2.4B) follows v1.
- Editing and control
- Text or audio style, steered live
- Licence
- v1 weights CC-BY 4.0; MRT2 weights licence not stated.
Royalty-free music APIs
Royalty-free music APIs
Background and production-music generators, mostly instrumental, often with stems and tempo control.
Beatoven.ai (Maestro)
Beatoven.ai
Background-music API that returns a mix plus four stems. Fairly Trained certified; says it trained on 3M+ tracks from rights holders.
- Editing and control
- Four stems (bass, chords, melody, percussion), looping flag
- Licence
- Commercial use advertised; end-user release terms not published.
Loudly
Loudly
Music API with explicit tempo, key and song-structure parameters; lossless WAV and stems.
- Editing and control
- BPM, key, duration, structure, stems
- Licence
- Royalty-free per vendor.
Mubert
Mubert
Generative-music API from 15 s to 25 min with tempo control and swappable part groups.
- Editing and control
- BPM, mood, intensity, part groups
- Licence
- Published sublicensing tiers; every plan bars release to streaming services and Content ID registration.
Soundraw
Soundraw
Production-music API trained, it says, only on music it owns. MP3, WAV and stem WAVs.
- Editing and control
- Stems, length, mood and genre
- Licence
- Commercial use advertised; licence agreement not public.
CassetteAI
CassetteAI
Fast instrumental generator: a 3-minute 44.1 kHz WAV in under 10 seconds.
- Editing and control
- Text only
- Licence
- Listed for commercial use on fal.
Rights and lawsuits
The legal
picture.
Who has settled, who is still in court, and the rules that apply to what you release.
- Dec 2, 2026End of the grace period for systems already on the EU market to meet the AI Act's Article 50 machine-readable marking duty, which applies from 2 Aug 2026. digital-strategy.ec.europa.eu
- Sep 18, 2026UMG and Sony file a second suit against Suno over 60,202 recordings, arguing its v6 models derive from earlier ones. Suno says v6 was trained from scratch without their recordings. musicbusinessworldwide.com
- Sep 10, 2026ElevenLabs signs a licensing deal with UMG. variety.com
- Jul 31, 2026Munich Regional Court I finds for GEMA against Suno on six works — training, memorisation, offering the model and outputs. Not final. twobirds.com
- Nov 2025Warner Music Group settles with Suno and Udio and licenses both; WMG also forms an alliance with Stability AI. musicbusinessworldwide.com
- Oct 2025UMG settles with Udio, which disables downloads the next day; UMG and Stability AI announce a strategic alliance. musicbusinessworldwide.com
- Jan 29, 2025US Copyright Office: purely AI-generated material is not copyrightable; human contribution is assessed case by case. copyright.gov
Method
How we
checked.
Each entry was checked on October 28, 2026 against the maker's own documentation, pricing, licence files and terms, with news coverage used only for lawsuits and deals. Where a fact could not be confirmed from a primary source it is left out or marked "not published". Lineups in this field change monthly — prices and versions here may already have moved.
In SoundGen, the Generate page lets you pick the model, and each option states its limits before you spend credits. What you may do with a track depends on the model that made it — see Terms §4.
Product names and trademarks belong to their respective owners. They are used here only to identify the products being described; SoundGen is not affiliated with, endorsed by or sponsored by any of these companies.
Spotted something out of date? Tell us and we'll correct it. For SoundGen against other music apps, see the comparison.
Generate, then
keep editing.
Your first 100 credits are on us — that's 5 thirty-second generations to start with. No card required.