Stable Audio

Where Stable Audio sits

Stable Audio is Stability AI's music and sound-effects generation family — a hosted web app at stableaudio.com, a DAW plugin, a developer API on the Stability platform, and, unusually for this category, downloadable model weights on Hugging Face. It launched in September 2023 as the audio sibling of Stable Diffusion, and the company has since applied the same open-weights philosophy it used to reshape image generation: three of the four Stable Audio 3.0 models released in May 2026 ship as weights you can run yourself.

That single fact is the reason Stable Audio remains on any shortlist. Suno, Udio, ElevenLabs Music and Google's Lyria are all closed products — your pipeline ends where their export policy begins. Suno capped downloads in September 2026 (20 per month on Pro, 60 on Premier, applied retroactively to existing libraries) and Udio turned downloads off entirely in October 2025 after its Universal deal. With Stable Audio, the output is a WAV file you own, and if the hosted product disappeared tomorrow, the open weights would still run on your own hardware. For anyone building game audio, video beds, podcast scoring or production music on a schedule, that changes the risk calculation.

The licensing story behind the model is the other differentiator. Stability published training-data attribution for 1,278,902 recordings, drawn from a licensed AudioSparx catalogue plus Creative Commons material from Freesound, and has announced partnerships with Warner Music Group and Universal Music Group for the 3.0 generation. In a category still contending with unlicensed-training lawsuits, Stable Audio is the commercially cautious pick — and the company says so at length in its own comparison pages.

What it actually does

You describe the sound you want — instruments, genre, mood, tempo, format — and it generates instrumental music, loops, or sound effects. Since version 2.0 it can also take an audio input and restyle it, and 2.5 added inpainting so you can regenerate one section of a track without re-rolling the whole thing. Track length is the practical headline: up to six minutes and twenty seconds on the 3.0 models, enough for a complete instrumental with intro, development and outro in a single pass rather than stitched clips with audible seams. Output is stereo 44.1 kHz; downloads come as a WAV mixdown or, since the app rebuild, separate stems in a ZIP — stem export is labelled experimental.

The model line is worth mapping because the pricing pages do not do it for you. Stable Audio 2.0 (April 2024) brought full-length structured tracks; Stable Audio Open 1.0 (June 2024) opened the weights but capped at roughly 47 seconds; Open Small (2025) shrank to 341M parameters and runs on phone-class Arm CPUs; Stable Audio 2.5 (September 2025) is the current hosted flagship with inpainting; and the Stable Audio 3.0 family (May 2026) splits into Small and Small SFX (459M, up to two minutes, on-device capable), Medium (1.4B, up to 6:20, LoRA fine-tuning) and Large (2.7B, API and enterprise deployments only). Small, Small SFX and Medium are open weights under Stability's Community License: free for commercial use below $1M in annual revenue, with an Enterprise licence required above that. Large is excluded from the self-host path.

Pricing, at time of writing, is where this product gets messy. The hosted app runs a credit-based subscription whose tiers have been renamed and re-priced more than once this year. Recent independent checks list the ladder roughly as: an entry tier around $12 per month with about 660 credits, mid tiers at $30 and $90, and a ceiling at $199 per month — up from $89.99 earlier in the year. Some listings still show $11.99 Pro / $29.99 Studio / $89.99 Max. Whether a standing free plan exists is genuinely unresolved: one recent review reports 50 credits a month on a free plan, another reports the free plan is no longer sold and new signups get 50 one-time credits. The API is clearer: credits at $0.01 each, 25 free on signup, with a Stable Audio 2.5 generation at 20 credits ($0.20) and a 3.0 generation at 26 ($0.26). Models also run through fal, Replicate and ComfyUI.

My testing notes

I have been generating with Stable Audio since the 2.0 release and came back for the 3.0 open-weight drop. Two observations.

First, the tool is genuinely strong at texture and weak at hooks — and it is designed that way. My electronic, ambient and cinematic prompts returned detailed pads, believable stereo image, and coherent structure across multi-minute renders; the six-minute ceiling means an ambient bed or a game-music loop comes out as one continuous piece rather than a loop I have to disguise. Acoustic and lo-fi prompts were the flakier end — fingerpicked guitar and brushed drums landed, but the mix felt flatter and occasional transitions arrive abruptly, the same complaint levelled at 2.0 in early comparisons. And the vocal question answers itself: the vendor's own prompt guide states the model is not intended to generate intelligible vocals, and my prompts bore that out. Lyrics in a prompt describe a sound; they are not sung. For a song with words, this is the wrong tool and Suno is the right one.

Second, the open weights are real, but pick the model with care. Medium needs a serious GPU or an Apple Silicon box; Small runs on CPU and is fine for sketches and sound effects but caps at two minutes and a colleague of mine described its output as sample-length rather than song-length — the residue of the older Open 1.0's 47-second ceiling. Running locally also means you own the compute bill instead of a subscription, which pencils out for teams with hardware already warm. The Community Licence terms are the cleanest in the category below $1M revenue; above it, budget for an Enterprise conversation.

What it does not do

No vocals, no lyrics, no voice generation — by design, not by tier gate. The FULL MIX and MULTI-TRACK buttons control stems, not singers. No hook-driven pop songwriting, no verse-chorus storytelling; the output is production music and sound design. The hosted app that ships today is labelled an early beta by the vendor, with its own published list of unfinished functions and stem export marked experimental.

The commercial documentation is the weak flank. The vendor's own pages name the same plans three different ways — one plan name maps to different price points on the pricing page and in the knowledge base — and published terms and help pages disagree on refund policy and on whether purchased credits expire. I could not resolve whether a standing free plan exists right now. Before you commit, get the current tier name, price, credit allowance and refund position in writing. Reliability of the hosted service is not something Stability publishes numbers for.

One more limit worth stating plainly: distribution platforms screen AI-generated audio regardless of which model made it, and no tool can promise what Spotify, YouTube or a distributor will do with your upload. The clean licensing reduces your legal exposure; it does not guarantee platform acceptance.

Verdict

Stable Audio is the production-music and sound-design tool of the category rather than a song machine: six-minute instrumental renders, audio-to-audio restyling, inpainting, stem export, a self-serve API at roughly 26 cents a generation, and open weights you can run and fine-tune under a licence that is free below $1M revenue. Rated 3.8 overall — core functionality is strong inside its instrumental lane and the open-weight path is unique among major players, value is decent but slipping as the free tier thins and the ceiling price doubled, usability is held back by a self-described beta app and complex tag-style prompting, the ecosystem (Hugging Face, ComfyUI, fal, Replicate, DAW plugin) is the best in the category, and reliability signals suffer from the vendor's own contradictory pages. If you need instrumentals, loops, SFX or a model you can self-host, it is the pick. If you need a singer, stop here.

Sources

• https://www.stability.ai/stable-audio
• https://stability.ai/explainers/stable-audio-vs-competitors-licensing-export-rights-and-self-hosting-compared
• https://platform.stability.ai/
• https://huggingface.co/stabilityai
• https://github.com/Stability-AI/stable-audio-tools
• https://arxiv.org/abs/2605.17991
• https://bigbangindex.com/tools/stable-audio
• https://theproofdesk.com/stability-ai-review
• https://aitoolsofthetrade.com/reviews/stable-audio
• https://amrytt.com/stable-audio-review
• https://undetectr.com/blog/stable-audio-review
• https://aitoolspolice.com/reviews/stable-audio/
• https://vibecontent.net/apps/stable-audio
• https://aiqora.io/tools/stable-audio
• https://decrypt.co/224729/stability-ai-launches-stable-audio-2-0-how-does-it-stack-up-against-the-mindblowing-suno-v3
• https://agentbrisk.com/agents/stable-audio