Stable Audio Review (2026): Pricing, Commercial Rights & the Verdict

Stable Audio is not the product this page described three weeks ago. Stability AI relaunched the web app on 18 August 2026, moved it onto a credit meter, shipped a DAW plugin alongside it, replaced the plan names, and left the old plan names standing on its own help pages. The prices below were read from the vendor's pricing page and documentation on 4 September 2026, and every figure names the document it came from.

Two things decide whether this tool fits before any question of sound quality. Stable Audio is built for instrumental music and sound effects, and its own prompt guide says the model is not intended to generate intelligible vocals. And the entry price is now $12 a month for 660 credits, with no free plan on sale, which changes the answer to "can I try this properly first" from yes to a 50-credit trial run.

Stable Audio scores 3.9 out of 5 on this pass, down from 4.1. The move is not about the sound. Sound design (4.4) and instrumental output quality (4.3) are unchanged, because nothing in this update touched them. What fell is licensing and pricing clarity, and value: the free plan is gone, the top tier went from $89.99 to $199, the vendor calls the new app an early beta and publishes its own list of what does not work yet, and its pages contradict each other in eight places that a buyer would want settled.

What Stable Audio is now: three routes under one name

Stable Audio is an AI music generator from Stability AI that produces original instrumental music and sound effects from text prompts, and edits existing audio. In catalog terms it is a SoftwareApplication published by Stability AI, the British AI company listed on Wikidata; the product itself has its own Wikidata entry. It is not a generic concept of "stability"; throughout this review, Stable Audio means the music and audio system, nothing else.

Since August 2026 the name covers three different things, and the pricing question has a different answer for each.

The relaunched web app and DAW plugin

The current hosted product is the rebuilt app at stableaudio.com plus a plugin that runs inside a digital audio workstation. The pricing page states it plainly at the top of the plan grid: "All plans come with full access to the Stable Audio web app and the Stable Audio DAW plugin." Credits are shared between the two, so a session in the plugin draws down the same balance as a session in the browser. The vendor's prompt guide for this app describes Stable Audio 3.0.

The vendor also calls it an early beta, and publishes a list at /docs/manual headed "What doesn't work yet". Two entries on it are worth knowing before you subscribe: "Re-recording a time range inside a track currently fails" and "Custom effect generation: Unavailable for now." A vendor that publishes its own defect list is doing something most do not, and the same documentation set says a failed generation is not charged, which is the rule that matters if you run into one.

The legacy app, still running for people who had it

Anyone who signed up before August 2026 has a second product available. Legacy subscribers keep their subscription, their old price, their existing single-track library and the old generator, reachable through a "Use Legacy Stable Audio" link. That route is closed to everyone else, and the section on legacy pricing below covers the one action that closes it for you too.

Stable Audio 3 open-weight models, for developers

Stable Audio 3, released on 20 May 2026, is a family of open-weight models: the weights themselves are published for download and modification instead of being reachable only through a website. The family has four members, a Small SFX model and a Small music model at 459M parameters each, a Medium model at 1.4B, and a Large model at 2.7B. Small and Medium are free to download and run yourself; Large is API-only. This route has its own licence, its own costs and its own revenue threshold, none of which are the subscription's, and the section further down keeps them apart.

Stable Audio is a narrow specialist inside a much larger field; our full library of AI tool reviews covers the tools built for other jobs.

How we reviewed this

Stability AI's own documents are the spine of this review: the pricing page, the credit-cost documentation at /docs, the knowledge base, the prompt guide, the plugin page, the manual's defect list, and the Terms of Service, all read on 4 September 2026. Alongside them sits our own dated record of the same product from 12 August 2026, captured with archive URLs while collecting data for our pricing transparency study, which is what makes it possible to say what changed and when, not only what is true today. No hands-on session was run, and nothing here is dressed up as one.

That basis matters in this category, because the pages currently ranking for Stable Audio tend to fall into two camps, and neither is a disinterested test. Several are thin directory listings that aggregate user reviews with no original analysis. The most technically detailed competing review is written by a disclosed AI agent, with deep spec coverage but no output evidence. Stable Audio is not a product this newsroom sells or competes with in any way. A claim about audio quality or generation behavior that only a controlled test could settle gets flagged instead of scored: per-genre output samples, a timed generation log, and a real plugin session are the specific artifacts a future hands-on pass would add, and none of them exist today.

That gap is closed as of 5 September 2026. This page previously carried no attributed, dated user reports at all, which was a standards failure of ours rather than an absence in the record, and it now carries fourteen dated comments with permalinks, every one of them read at source before it was quoted, set out below. One caveat travels with them and is repeated where they appear: they come from a single platform, because Reddit could not be searched on that pass and no listing was reachable on G2, Trustpilot or Capterra. Where this review still describes how the tool is received beyond those reports, it is reading the vendor's own positioning and the model paper's stated scope.

Disclosure

AI Tools Police earns affiliate commissions when readers sign up for some tools we cover, which may include this one. That never changes a score, a documented figure, or whether we surface a weakness. This update lowers the score and withdraws claims that made the product look simpler to buy than it is. Stability AI has no stake in this newsroom, financial or otherwise, and this newsroom has none in Stability AI beyond the referral link disclosed above.

The rubric: how we score Stable Audio

Every one of the five scores below is defined first, so a 4.4 means the same thing whether a reader stops at this paragraph or skips straight to the table.

Instrumental output quality (4.3) measures how convincing the generated music sounds in Stable Audio's core territory: electronic, ambient, cinematic, and lo-fi instrumentals. This is the tool's home ground. Unchanged this pass, because nothing in the September document read bears on output quality.

Sound design and SFX (4.4) is the highest axis, and it is where Stable Audio leads vocal-first rivals. The model was built to generate sound effects and to edit audio, so foley, textures, and transitions are first-class outputs. Also unchanged.

Licensing and training-data clarity (3.8, was 4.2) measures how clean and well-documented the rights story is. The training-data half is unchanged and remains the strongest in the category. The documentation half is what fell: the licence matrix that used to state commercial rights per tier has been removed, the terms carry two different effective dates, and the refund and credit-expiry positions disagree between documents.

Value for money (3.6, was 3.9) weighs price against output. The entry price barely moved, from $11.99 to $12, but the free plan is no longer sold, the ceiling tier went from $89.99 to $199, and the product you are buying is one the vendor calls an early beta with a published list of broken functions.

Vocal capability (1.5) is low on purpose and is not a knock on the engineering. Stable Audio is built for instrumental work. If sung lyrics are the deliverable, this axis tells you to stop reading and open Suno instead.

Stable Audio pricing in 2026: the credit ladder

Stable Audio now meters generation in credits. Four plans are sold, all monthly, with no annual option and no toggle on the page. The figures below come from stableaudio.com/pricing, read 4 September 2026.

The credit counts and prices are the vendor's. The two right-hand columns are our arithmetic on the vendor's published rates, not figures Stability AI prints anywhere.

One more piece of arithmetic worth having before you pick a tier: the price per credit falls from about 1.8 cents on Solo to about 1.4 cents on Studio, so the top plan is roughly 22 percent cheaper per credit than the entry plan. That is a shallower volume discount than most credit-metered creative tools offer, which means there is little to gain from buying ahead of your actual monthly usage.

Our archived capture of this page from 12 August 2026 recorded a line reading "excluding VAT/sales taxes" under each paid price. The current page carries no tax note at all. Absence of a note is not a statement that no tax is added, so treat the displayed figure as pre-tax until the vendor says otherwise.

What a credit buys, and what costs nothing

The vendor's documentation at /docs prices each action, and this is the part most competing coverage skips. A FULL MIX costs 4 credits. A MULTI-TRACK costs 6, and the documentation is specific that this is charged "once, however many tracks come back", so a six-track return costs the same as a two-track return. Adding, re-cutting or extending costs 4 credits each time.

What costs nothing is the more useful half of that page. Playback, effects, edits, printing, bouncing, mixdown, stems export and sharing are all documented as free and unlimited. The meter runs when the model generates and stops when you work on what it gave you, which is a materially different cost shape from tools that charge for exports or downloads. The documentation also states that "A generation that fails is not charged." Our August record of this product had to code that question as unknown, because no document then addressed it; the vendor answering it in print is a small, real improvement in what a buyer can know before paying.

What $12 actually buys

At the entry tier, $12 a month buys 165 full-mix starts or 110 multi-track starts. Freelancers think in deliverables, so the useful number is the cost of one track you would actually ship: generate ten keepers in a month and each cost about $1.20, generate forty and each cost about $0.30. The generous no-charge rules push in your favour here, because the re-cuts and mixdown passes that turn a start into a keeper are either free or cost the same 4 credits as a fresh mix.

A like-for-like comparison with the old ladder is not available, and we are not going to fake one. Our 12 August record shows the entry plan at $11.99 for 250 "track generations", which works out at about 4.8 cents per generation. Today's entry plan works out at about 7.3 cents per full-mix start. The units are close but not the same thing, since the old app metered generations and a separate monthly upload allowance in minutes, while the new one meters credits across a wider set of actions and gives away the editing. Read the two numbers as a direction of travel, not as a measurement.

Is there a free plan?

No free plan card appears on the pricing page, and no trial appears in any document we read. What exists is a one-time grant, stated in the vendor's knowledge base: "Every new account receives 50 free credits upon signup to test out the platform." At the published rates that is 12 full-mix generations, or 8 multi-track generations, in total. Not per month.

Twelve generations is enough to learn whether the model handles your genre and your prompt style, and it is not enough to finish anything. That is the wall this product now has, and it arrives faster than most people expect: a single scoring brief that needs three moods explored at two lengths each is already past it. The honest read is that Solo at $12 is the real entry point, and the 50 credits are a look at the interface, not a working free tier. If you need to evaluate properly before paying, the open-weight Stable Audio 3 Small and Medium models are free to download and will tell you what the model family sounds like without a card, though they are not the same build as the hosted app and come with their own licence to read.

One thing this page will not tell you either way: whether accounts created before the relaunch keep a standing free plan. Our own record from 12 August documents a perpetual free plan with 10 monthly track generations, and the current pricing page's FAQ still says "The free plan renews on the first of each month" while selling no free plan. Whether that sentence describes a live entitlement for old accounts or is simply a page nobody updated, the vendor's documents do not settle, and guessing would be worse than saying so.

Legacy subscribers: what you keep, and what one plan change costs you

Anyone who subscribed before August 2026 is in a different position from a new buyer, and this is the decision most likely to cost a reader money this month. Legacy subscribers keep their subscription and their price, keep their existing single-track library, keep access to the old generator through a "Use Legacy Stable Audio" link, and are granted credits on the new platform on top: $11.99 receives 660 credits, $29.99 receives 1,800, and $89.99 receives 6,000.

Line those grants up against the new ladder and the arithmetic is stark. A legacy $11.99 subscriber receives exactly what a new Solo subscriber pays $12 for, plus the old app, plus the old library. The same holds at $29.99 against Session and at $89.99 against Producer. Legacy is the better deal at every price point where it exists.

Then read the vendor's own sentence: "Legacy subscription plans are no longer available for new purchases or plan changes." That covers plan changes, not just new sign-ups, so moving between plans is a one-way door. A legacy subscriber who upgrades or downgrades once lands on the current ladder permanently and cannot go back. Note also that legacy tops out at the equivalent of Producer, since the old ladder had no $199 tier, so a legacy subscriber who needs 14,000 credits a month has to give up legacy pricing to get them. Anyone in that position should price the whole year both ways before touching the plan selector.

Where the vendor's own documents disagree

Eight disagreements sit in Stability AI's published material, all read on 4 September 2026. A company that rebuilds a product in a few weeks will leave old pages behind, and none of this is exotic. The reason it belongs in a review is that several of these decide what you pay, what you can get back, and which product you are actually buying.

1. A free plan that is not sold but renews. The pricing page offers no free plan, and its own FAQ on the same page says "The free plan renews on the first of each month." Those two statements sit within one scroll of each other. Anyone who arrives to claim the free plan that FAQ describes will find nothing to claim.

2. Three sets of plan names, and Studio means two different prices. The pricing page sells Solo, Session, Producer and Studio, where Studio is $199. The knowledge base calls the same four Creator, Pro, Studio and Max, where Studio is $90. The pricing FAQ refers to "Pro and higher tiers", a name the pricing page does not sell at all. A reader paying $90 a month is on Producer according to one vendor page and Studio according to another. Buy by price and credit count.

3. Purchased credits expire, or never do. The knowledge base states that "Purchased credit packs never expire". Section 7d of the Terms of Service states that "Unused service credits ... expire one year after the date of purchase or issuance." Both are current. What happens to unused monthly subscription credits at the end of a billing cycle is not settled by the documents we read, in either direction.

4. Refunds are refused and offered. Section 7a of the Terms of Service says "All payments are non-refundable, except where required by law". The pricing FAQ and the vendor's refund article publish a discretionary 48-hour refund process. Our 12 August record captured the same disagreement, six days before the relaunch, so it is not something the rebuild introduced.

5. The terms carry two effective dates. stableaudio.com/terms says the terms are "Effective July 31th, 2025", the typo included. The document it links to is headed "Effective: September 30, 2026", a date 26 days in the future on the day we read it. We report both and adjudicate neither.

6. Pro Tools works, or is coming soon. The homepage says the plugin works with Pro Tools. The plugin page says AAX support, which is the format Pro Tools uses, is "coming soon". If you are a Pro Tools user, confirm before you subscribe.

7. Six minutes, or three. The homepage and the launch post say tracks run up to six minutes. The app's length picker offers up to 3:00, with EXTEND as the documented route past it. Extending costs 4 credits each time, so the longer number is reachable but not free.

8. The help pages document an older build. The knowledge base still shows a "FULL MIX / STEMS" toggle and calls the app "ST-1", where the live app shows "FULL MIX / MULTI-TRACK". Two URLs our August record cited have since been redirected: stableaudio.com/faqs now goes to the homepage, and /user-guide now goes to /docs. The archived captures remain, which is why we can still say what those pages said.

Commercial rights: what the current documents actually say

Commercial rights are the decision gate for most readers of this review, and the honest answer changed with the relaunch. Nothing Stability AI publishes today conditions commercial use on a subscription tier. The words "commercial", "tier" and "plan" do not appear in the current Terms of Service at all. Output ownership is addressed directly, in section 4a: "we assign to you all of our right, title, and interest (if any) in the Outputs. So as between Stability and you, you own the Output (to the extent permitted by applicable law)."

That is a stronger position for a paying user than the one this page described in June, when commercial use was gated behind a paid plan and capped by an audience ceiling. It is also a less specific one. An assignment of output rights in the terms is not the same instrument as a licence matrix that lists which uses each plan permits, and the matrix is what has gone.

The $1,000,000 threshold belongs to a different document

A $1,000,000 annual-revenue threshold circulates in coverage of Stable Audio, and it is real. It is a term of Stability AI's Community License, which governs the downloadable model weights and self-hosting, and above that revenue level an enterprise licence applies instead. It is not a term of the hosted subscription, and it does not limit what a Solo or Studio subscriber may do with a track generated in the web app. Stability AI's own blog writing blurs the two documents. If you subscribe, the Terms of Service is your instrument. If you download weights, the Community License is yours. Read the one that applies to you, and if you are doing both, read both.

The 100,000-MAU term we published, and its withdrawal

Until this update, this page said commercial use ran up to a ceiling of 100,000 monthly active users. That was accurate when we wrote it. It came from a licence matrix on the old pricing page, which our own research desk recorded and archived on 12 August 2026 while collecting the Pricing Transparency Audit. The matrix granted commercial project and commercial music release rights from the Creator licence upward, and reserved three rows for the sales-gated Enterprise licence: commercial products above 100,000 monthly active users, film, television and advertising, and apps and games.

That matrix has been removed. The figure is a withdrawn term, not an error we are retracting, and the distinction matters for anyone who signed a client brief against it. Two consequences follow. Nothing published today reinstates the ceiling, so a subscriber has no current MAU cap to breach. And the two other Enterprise-only rows, film, television and advertising, and apps and games, were never mentioned on this page even while the matrix was live, which was our omission: those carve-outs were more likely to bite a working freelancer than the MAU row we did print. No current document restricts either use.

Training data and the AudioSparx opt-out lawsuit

Stable Audio's strongest trust argument is its training data, and it deserves an honest accounting instead of a marketing line. Stability AI documents that Stable Audio was trained on 1,278,902 recordings sourced from AudioSparx and Freesound under CC-0, CC-BY, and CC-Sampling+ terms, with the breakdown published in its attribution materials. Set against the unresolved copyright litigation hanging over vocal-first rivals, a fully attributed, licensed corpus is a real advantage for anyone who needs to defend where their music came from.

The honest complication is that the licensing is not entirely settled. In December 2025, artist Anders Manga reportedly brought a lawsuit concerning the AudioSparx opt-out process, raising questions about whether some AudioSparx-licensed material was used despite artist objections. The suit is reported as ongoing, and we did not re-check its status in this pass, so treat that as of the date it was last verified on this page and not as of today. It does not negate Stable Audio's documented provenance, and it is narrower than the broad infringement cases facing other tools, but it is a live uncertainty worth tracking before you build a catalog on this tool.

Distribution: the AI disclosure step still applies

Holding the rights to your output is necessary but not sufficient to get AI music onto streaming platforms. The workflow is: generate, export, then submit through a distributor such as DistroKid, TuneCore, or CD Baby. At that submission step the distributors require AI disclosure, and the platforms screen uploads independently of whatever rights you hold. Your terms with Stability AI settle whether you may use the music. They do not exempt you from a distributor's disclosure rules or a platform's screening. Expect to declare AI involvement, and set expectations for any release plan accordingly.

Vocals: what the vendor says, and where its own pages disagree

Stable Audio does not produce sung lyrics, and this page has said so since June. The mechanism it gave was wrong. There has never been a tier gate on vocals at any price, so no plan unlocks them and none ever did. The limit is in what the model was built and trained to do.

The vendor's prompt guide states it directly: "Stable Audio 3.0 is designed primarily for instrumental work. It can sometimes create non-lexical vocal-like textures, but it is not intended to generate intelligible vocals." Stability AI's own research paper on the model family, arXiv:2605.17991, describes the models as built for "instrumental music and sound effects generation" and documents training on prompts marked "VocalType: Instrumental". Two independent vendor sources, saying the same thing.

Two other vendor pages pull the other way. A knowledge-base article offers "+ STEM: Generate additional layered instruments or vocal stems", and a moderation article uses an example prompt naming "a breathy female vocal". We have no basis for deciding which of these four pages describes the shipping model, so we report the disagreement and leave it open. What we can rule out is a common misreading of the interface: the FULL MIX and MULTI-TRACK buttons are a stems-versus-single-file control, not a vocals switch, and choosing MULTI-TRACK will not produce a lead vocal.

The practical guidance is unchanged. If your deliverable is a song with sung lyrics, Suno is the right tool and this one is not, whatever the moderation article's example prompt implies.

Sound quality by genre

Stable Audio is positioned by its maker as a producer's instrumental and sound-design tool, not a one-click song machine, and the model paper's stated scope agrees. The notes below reflect that documented scope and the tool's own positioning rather than a controlled measurement; treat them as a map of where the tool is built to be strong, and verify against your own prompts. Since 5 September 2026 they are also corroborated from outside the vendor: the user reports below independently reach the same electronic-strong, acoustic-weak split across three model generations.

Electronic and EDM is the lane the model is most clearly built for. The 44.1 kHz stereo output and the training emphasis on structured, beat-driven material suit coherent loops and builds. Ambient and cinematic is the other fit: pads, drones, and evolving textures suit a model designed for sound design, and track length matters here, because cinematic beds and background scores need duration. Lo-fi and acoustic is the weaker end, the lane a vocal-first competitor's broader genre training is built to cover and this model's instrumental emphasis is not. Across all of it, the caveat to carry is that this is raw material for a project, strong enough to drop into a video or a game scene, and not a finished, mixed master.

What real users report

This page carried no attributed user evidence until 5 September 2026, which was a gap in our standards rather than in the record, and this section closes it. Every report below names the account, gives the date, and links to the exact comment.

One limit belongs at the top, because it shapes everything under it. All of this comes from a single platform. Hacker News is where Stable Audio's releases were actually discussed at length, and the threads run the product's whole public life, from the September 2023 launch to the Stable Audio 3 release in May 2026. But Reddit could not be searched at all on this pass, and G2, Trustpilot and Capterra have no listing for the product or would not serve one. So this is a developer-heavy and release-day-heavy sample, not a cross section of everyone who uses the tool. Read it as the record of one specific room.

The genre split is the most consistent finding, and it holds across three years. At launch cwillu (September 13, 2023) singled out a rock drum loop as the credible output while dismissing most of the rest. Five months later bane (February 13, 2024) put a mechanism under it: the rave material "sounds great" because electronic music tolerates unusual construction, while "the guitar sounds weird because it doesn't sound like chords a human hand can make." Two years after that, on Stable Audio 3's release day, rafadc (May 20, 2026) reached the same place from a self-hosted run: "It is better for electronica than for any other genre." Three independent accounts, three model generations, one conclusion, and it matches what the vendor's own positioning claims for the tool.

The second pattern is that output is raw material, not a finished master. bufferoverflow (June 5, 2024) described "something unpleasant about its high frequencies" and no vocal capability. hununu (May 20, 2026) judged the vendor's own website sample usable "as sample materials for song making but definitely lacking frequency range expected today as a final product." rafadc, above, called the same release "insanely fast" at under two seconds for two minutes of audio on a consumer GPU, and in the same breath said it "sounds too much like general midi." Fast and usable as a component, weaker as an end product, from people running it themselves.

On licensing, the picture is genuinely mixed and the two halves point opposite ways. evmaki (May 20, 2026), a developer who had the open weights running in their own instrument the morning of release and who discloses building a commercial product on them, praised "the attention to detail with licensing in the training set." Against that, simonw (June 5, 2024) objected to the open-weight release being called open source when it ships under the Stability AI Non-Commercial Research Community License, and called Stability "one of the worse offenders for abusing the term." Both are about licensing and neither refutes the other: the training data and the model license are separate questions, and the vendor has a defensible answer on one and a contested label on the other.

One reported surprise matches a defect this review found separately. nullandvoid (February 13, 2024) went to generate sound effects for a game and hit an enterprise-license requirement, asking why it did not fall under the page's own "commercial products below 100,000 MAU" allowance. AuryGlenz the same day: "Too bad according to their page you need an enterprise license for even indie games." That is the same carve-out shape our own 4 September pass found we had omitted from this page, reported by buyers two and a half years earlier.

A dated quotation worth keeping for what it shows about change. On April 3, 2024, livrem quoted the then-current Terms as tier-gating commercial use, with paid tiers permitted "some (but not all) commercial uses" and Basic users encouraged toward non-commercial work. The Terms we read on 4 September 2026 carry no such tier-gating. We are relying on a reader's quotation of a document we did not retrieve ourselves in its 2024 state, so treat the exact wording as reported rather than verified. What it does establish is that the commercial-rights structure changed at some point nobody has dated, which is the honest state of that question and the reason the commercial-rights section on this page is written the way it is.

Finally, the reliability reports. coldcode hit a 500 error on the Stable Audio 2.0 launch in April 2024, and 0gs found stableaudio.com unreachable on Stable Audio 3's release day in May 2026. Two reports two years apart is a coincidence rather than a pattern, and it is recorded here as such.

The resignation story that circulates about this product is backwards

It is worth correcting a claim readers will meet elsewhere, including in the Hacker News thread this section draws on, because it runs in the vendor's favour and we would print it either way.

shon (February 13, 2024) wrote that Ed Newton-Rex, hired to build Stable Audio, "quit shortly after it was released due to concerns around copyright and the training data being used." Newton-Rex did resign, on November 15, 2023, as Stability AI's VP of Audio. The reason was not this product's training data. In his own published resignation letter he gives it plainly: he disagreed with the company's position, stated in Stability's submission to the US Copyright Office, that training generative AI on copyrighted works is fair use. In the same letter he singles out the audio team's product as the thing he was proud of, "a state-of-the-art AI music generation product trained on licensed training data, sharing the revenue from the model with rights-holders," and adds that he thinks Stability takes a more nuanced view than some competitors.

So the person best placed to know, and with every reason to say otherwise on his way out, is on record that Stable Audio's own training data was licensed and revenue-shared. Set that beside then-CEO Emad Mostaque's own comment on the same Hacker News thread, emadm (February 13, 2024), "stableaudio.com is fully licensed," and the vendor's claim is one of the few in this category with independent support behind it. The contested label, per simonw above, is on the separate open-weight release, not on the hosted product's corpus. Keeping those two apart is the whole difference between a fair criticism and a repeated one.

Features: text-to-music, sound effects, and audio editing

Stable Audio's feature set is built around three verbs: generate, edit, and, for developers, fine-tune. The first two live in the hosted app and its plugin.

Track length: six minutes or three?

Track length is the one headline spec the vendor states two ways. The homepage and the launch post say up to six minutes. The app's length picker offers up to 3:00, and EXTEND is the documented route past that, at 4 credits per extension. Both numbers are the vendor's, and until it reconciles them the safe planning assumption is a three-minute generation that you lengthen deliberately, at 4 credits a step. This page previously printed the six-minute figure as a straight capability, which was reading the marketing page and not the app.

Sound effects

A genuine differentiator is that Stable Audio is also a sound effects generator. The same prompt box that makes music makes foley and textures: rain on a window, a sci-fi door, a crowd murmur. The dedicated Small SFX open-weight model exists specifically for this, and it is a use case that vocal-first tools like Suno do not target at all. For video editors and game developers who need a quick, licensable sound effect, this is the feature that justifies a look.

Editing, stems and the DAW plugin

Editing is where the relaunch changed this review's conclusion. The old hosted app returned a stereo mix and nothing else, and this page said so: no separated stem export, import the mix into your workstation and arrange around it. That is no longer what the documentation describes. The current app offers MULTI-TRACK generation, charged 6 credits once however many tracks return, and lists stems export among the actions that cost no credits at all, alongside mixdown, bouncing and printing. A producer who needs separate layers to work with now has a documented route to them.

Two caveats belong with that. The knowledge base still documents the older build's "FULL MIX / STEMS" toggle instead of the live "FULL MIX / MULTI-TRACK" control, so screenshots in the help pages will not match what you see. And the plugin's workstation support is stated inconsistently: the homepage names Pro Tools, while the plugin page lists AAX, the format Pro Tools requires, as "coming soon". Confirm your own workstation is supported before subscribing on the strength of the plugin.

Audio-to-audio editing remains: you can upload a reference recording and regenerate a section of it instead of only generating from a blank prompt. Against that, the vendor's own defect list says re-recording a time range inside a track currently fails, so the feature and its published limits should be read together.

Stable Audio 3 open-weight option, for developers

The developer product is a set of downloadable models, and choosing it is a different decision from subscribing to the app.

Open-weight means the model weights are published so you can download, run, and modify them, which is distinct from "open-source", a statement about the licensing of code and weights. Stable Audio 3's Small and Medium models are open-weight and free to obtain; the Large model is reachable only through Stability AI's API. Developers can apply LoRA fine-tuning, a method of adapting a model to a specific sound or style by training a small add-on instead of retraining the whole network, to the open variants.

Stability AI distributes these models under its Community License, a separate instrument from the subscription terms. That licence is more generous on output ownership than most readers assume: it states that "you own outputs generated from the Core Models or Derivative Works (such as fine-tunes) and therefore can use those outputs at your discretion, so long as you do so in compliance with applicable law." The free-use threshold Stability AI publishes for its Core Models sits at commercial use under roughly USD $1M in trailing annual revenue, above which an enterprise licence applies. What the public licensing page does not spell out is whether Stable Audio 3 specifically is classified under that same Core Model threshold or carries model-specific terms, so a team planning to self-host at meaningful commercial scale should get that classification confirmed in writing instead of assuming the general policy carries over.

SA3 Small vs Medium vs Large: model comparison

The table below is the access map that decides whether the open-weight route works for your hardware and your quality bar.

The practical wall is in the last row. If your pipeline needs the best-sounding model, you cannot self-host it; you call the API, which is a different cost and a different dependency from running weights on your own machine. The Small and Medium models are genuinely accessible, and a Small model runs on a capable laptop, but a developer who downloads them expecting Large-tier quality will be disappointed. Name the tier you need before you plan the build.

Stable Audio vs Suno vs Udio

The comparison most readers want is short: which of the three tools fits which job. The row that decides it is vocals.

Competitor figures are as last verified on their own review pages and are not part of the 4 September document read.

The decision rule is clean. Need a finished song with vocals? Suno every time, and our Suno review covers it. Want audiophile-grade production with a settled major-label licence? Udio, covered in our Udio review. Need licensing-clean instrumentals, sound effects, editable multi-track output, or a model you can self-host? Stable Audio. One note on the free-tier column: Stable Audio is now the only one of the three with no standing free allowance, which changes how you evaluate them side by side, since you can keep testing the other two for nothing.

Licensing safety vs Suno and Udio

On the narrow question of training-data risk, the three tools sit in a row. Suno settled with Warner Music Group but still faces Universal and Sony litigation, so its training-data question is not closed. Udio relaunched with a Universal Music Group licence after an October 2025 settlement, which resolved its biggest exposure, though Sony and Warner suits against it are still reported as open. Stable Audio built on documented, attributed licensing from the start, with the one open AudioSparx dispute noted above. For a commercial user ranking by how defensible the source of the music is, Stable Audio and the relaunched Udio lead, with Suno's exposure the largest of the three until its remaining cases resolve.

What changed, and what we got wrong

This page carried the old pricing for longer than it should have, so here is the record, dated.

The relaunch is not dated by the vendor. Stability AI's dated posts put the new app at 18 August 2026, and our own archived capture shows the old ladder still live on 12 August. The 18 August announcement does not mention pricing at all, there is no changelog, and the pricing page carries no effective date. So the change is not vendor-dated beyond that six-day window, and any page claiming a precise repricing date is guessing.

The withdrawn ladder. Until today this page listed a free plan at roughly 10 credits a month with no commercial rights, a Creator plan at about $11.99, and a Studio plan at about $29.99. None of those are on sale. The current ladder is Solo $12, Session $30, Producer $90 and Studio $199, all on credits. The old Studio figure is the one most likely to have misled a reader, because a plan called Studio still exists and now costs $199.

Our own two records disagreed, and the study was right. Our pricing transparency study recorded this product on 12 August 2026 and named the $11.99 plan Pro, quoting the vendor's own FAQ: "Can I use the music in commercial projects? Yes, if you're on the Pro subscription tier." The same archived record shows Creator was the name of the licence the paid tiers held, in the matrix beside Personal and Enterprise. This review printed the licence name as the plan name and called the $11.99 plan Creator. The study is right, the review was wrong, and the naming is fixed here instead of being left as a disagreement between two of our own pages.

The 100,000-MAU ceiling. Withdrawn by the vendor, not retracted by us. Covered in full in the commercial rights section above, along with the two Enterprise-only carve-outs we should have printed at the time and did not.

Stem export. This page said the hosted product offered no separated stem export. That was true of the app it described, and it is not true of the app that replaced it, which documents MULTI-TRACK generation and a no-cost stems export.

Vocals. The conclusion held, the mechanism did not. There was never a tier gate on vocals, and the phrase "no vocals at any tier" has been replaced with what the vendor's prompt guide and paper actually say, and with the two vendor pages that contradict them.

Sources that moved. Two URLs our August record cited now redirect: stableaudio.com/faqs to the homepage, and /user-guide to /docs. The archived captures behind that record are what allow the comparisons on this page.

Who should use Stable Audio, and who should skip it

Use Stable Audio if you are a freelance producer or content creator who needs licensing-clean instrumental tracks, cinematic and ambient beds, or sound effects, and you value a documented training-data story over a copyright gamble. Use it if you need royalty-safe background music for video, ads, or podcasts, because that is where its instrumental strength and its output-ownership terms line up. Use it if you are a developer who wants to self-host or fine-tune an open-weight audio model, with the caveat that the best variant is API-only. And if you are already a legacy subscriber, the strongest reason to stay is the one the vendor buried in a support article: your grandfathered price plus new-platform credits is better than anything on sale today.

Skip Stable Audio if you need vocals, in which case Suno is the tool. Skip it if you need to evaluate at length before paying, because 50 one-time credits is 12 generations and there is no free plan behind them. Skip it, for now, if you cannot work around an early-beta product whose maker publishes a list of functions that do not work yet. If Stable Audio is close but not quite right, our ranking of AI music generators weighs the alternatives for instrumental and royalty-safe work.

Common problems and fixes

A few recurring friction points have straightforward answers.

"I ran out of credits in a day." A full mix costs 4 credits and every re-cut or extension costs 4 more, so an exploratory session burns the 50-credit signup grant in about twelve generations. Editing, mixdown and stems export cost nothing, so the way to stretch a balance is to take a promising mix further with the free tools instead of re-rolling for a better start.

"The plan I bought has a different name in the help pages." Expected. The pricing page sells Solo, Session, Producer and Studio; the knowledge base calls them Creator, Pro, Studio and Max. The credit count and the price identify your plan reliably. The name does not.

"I'm on an old plan and want more credits." Check the legacy grants first. Changing plan is what ends legacy pricing, permanently, because the vendor closed legacy plans to plan changes as well as to new purchases.

"It won't generate vocals no matter how I prompt it." The vendor's prompt guide says the model is not intended to generate intelligible vocals, and the MULTI-TRACK button is a stems control, not a vocals switch. Use Suno for sung lyrics.

"I want to run the best model locally and can't find the weights." The Stable Audio 3 Large model is API-only; only Small and Medium are downloadable. Plan around the API if you need Large-tier quality.

Verdict

Stable Audio earns a 3.9 out of 5, down from 4.1, and the specialty is worth being clear about. It is the tool to reach for when you need licensing-clean instrumentals, sound effects, and editable multi-track output, backed by the most transparent training-data record in the category. Sound design (4.4) and instrumental quality (4.3) are unchanged, the terms now assign output ownership to you with no tier condition attached, and the open-weight models open a self-hosting door that neither Suno nor Udio offers.

What pulled the score down is everything around the product. The free plan is gone, replaced by 50 credits at signup. The ceiling tier went from $89.99 to $199. The vendor calls the app an early beta and publishes its own list of broken functions. And its pages disagree with each other about the plan names, the refund policy, whether credits expire, whether the terms took effect last year or take effect this month, and how long a track can be. None of that makes the audio worse. All of it makes the purchase harder to get right, and a buyer who reads only the pricing page will get at least one thing wrong. For producers and content creators in its lane, with the legacy trap and the plan names understood, it remains a straightforward recommendation. For anyone wanting vocals, it is still the wrong door.

Pricing and licensing on this page were read from the vendor's own documents on 4 September 2026 and are re-checked monthly. When live-account testing is added, this review will gain per-genre output samples, a timed generation log, and a plugin session in a real workstation; until then the assessment above rests on documented capability, the published terms, and our own dated archive of what those documents said before.