AI Stem Separation Pricing: The Costs and the Catches
LALAL.AI's annual plans cost exactly twelve times the monthly price. Its unlimited queue is free on every paid tier. The best-scoring option costs nothing.
No affiliate links on this page, and nothing here earns us anything. Prices were checked on 7 November 2026 against the sources listed at the end. How we know this.
The highest published separation score of anything here belongs to the tool that costs nothing. Demucs v4 reports 9.00 dB SDR on the MUSDB HQ test set, it’s free, and it runs on your own machine. The catch is a GPU requirement and a command line, and for plenty of people that catch settles it. Which is why cloud services can charge for the same job.
So the useful question isn’t which one is best. It’s what the paid services actually charge once you read the pricing page properly. That turns out to contain two surprises.
The short version
- The free option, Demucs, has the highest published quality score here. It needs a GPU or patience.
- LALAL.AI’s annual plans cost exactly twelve times the monthly price. There is no discount.
- The unlimited queue is free on both paid tiers. Only the fast queue is metered.
- Fast minutes do not roll over, which quietly doubles your real cost per minute if you underuse them.
LALAL.AI: the metered thing isn’t what you think
LALAL.AI runs two queues, and only one of them is rationed. Here’s the structure as published on 7 November 2026.
| Plan | Price | Relaxed Queue | Fast Queue | Upload limit |
|---|---|---|---|---|
| Starter | Free | 10 minutes | none | 200 MB per file |
| Lite | €6.75/mo, or €81 billed annually | Unlimited | 90 min/month | 2 GB per file |
| Pro | €13.50/mo, or €162 billed annually | Unlimited | 250 min/month | 2 GB per file |
On both paid plans the Relaxed Queue is unlimited. The monthly minute allowance only applies to the Fast Queue. Nearly every comparison we read treats that Fast Queue figure as the plan’s capacity, which gets the product backwards. On a €6.75 plan you can separate as much audio as you like. What the extra €6.75 buys you at Pro isn’t more separation, it’s less waiting.
What each paid plan actually gives you, by queue
That changes the decision. Working through an archive on your own schedule? Lite isn’t a 90 minute plan, it’s an unlimited plan with 90 minutes of queue skipping attached. Turning stems around against a deadline? The Fast minutes are the whole product, and you should price them that way.
Annual billing saves you nothing
Check the arithmetic on those annual figures.
- Lite: €81 divided by 12 is €6.75, identical to the monthly price.
- Pro: €162 divided by 12 is €13.50, identical to the monthly price.
There’s no annual discount. Paying twelve months up front gets you exactly the same rate as paying monthly, with your money handed over earlier, and the pricing page FAQ says subscriptions are non-refundable. We checked twice, because the usual pattern in this industry is two months free, and its absence is easy to assume away.
Unless you specifically want to be locked in, monthly is simply better. It’s rare to be able to say that about a SaaS pricing page without qualifying it.
What a Fast minute costs
| What you buy | Price | Fast minutes | Cost per minute |
|---|---|---|---|
| Lite subscription | €6.75/mo | 90/month | €0.075 |
| Pro subscription | €13.50/mo | 250/month | €0.054 |
| Master top-up | $50 one-off | 750 | $0.067 |
| Premium top-up | $190 one-off | 3,000 | $0.063 |
| Enterprise top-up | $300 one-off | 5,000 | $0.060 |
Subscription prices were shown to us in euros and the top-ups in US dollars, so the two halves of that table aren’t directly comparable without a conversion we aren’t going to invent. LALAL.AI also appears to localise pricing, so your figures may differ from ours. The per-minute column is the part worth carrying away either way.
Two things fall out of it. Within the subscriptions, Pro’s Fast minutes cost about 28% less each than Lite’s. The marginal maths is better still: the extra €6.75 between the tiers buys 160 more Fast minutes, which works out at roughly €0.042 per extra minute, about 44% below Lite’s rate. If you reliably go past 90 Fast minutes a month, upgrading is good value.
What a Pro Fast minute really costs, by how much of the allowance you use
The counterweight is that unused Fast minutes don’t roll over. The balance resets each billing period. So those per-minute figures are a best case that assumes you burn the whole allowance. Use half your Pro minutes and you paid €0.108 each, which is worse than Lite. Pay for the tier you actually use, not the one you mean to.
The one-off top-ups don’t expire with a billing cycle, and they scale mildly: the largest pack is about 10% cheaper per minute than the smallest. That’s a thin reward for a six-fold increase in commitment.
The free route, and what it really costs
Demucs is the reference implementation most of this field gets measured against. Its documentation reports 9.00 dB SDR on the MUSDB HQ test set for the fine-tuned v4 Hybrid Transformer model, separating drums, bass, vocals and other, with a six-stem variant adding guitar and piano. The project’s own notes say the piano performs poorly.
Ultimate Vocal Remover GUI is the usual way to avoid the command line. It’s a free front end bundling MDX-Net, VR Architecture and Demucs v3 and v4 models together, so in practice you can run several separators and keep whichever handles your material best.
The hardware requirement is the real cost.
Demucs needs Python 3.10 or newer, roughly 3 GB of GPU memory as a minimum and about 7 GB with default settings. Without a suitable GPU it runs on CPU, where the documentation estimates processing takes around 1.5 times the track’s duration.
UVR lists an Nvidia GTX 1060 6 GB as the minimum for GPU conversions, with 8 GB or more of VRAM recommended. It’s 64-bit only, on Windows 10 and up, macOS Big Sur and up, or a Debian or Arch based Linux install. AMD support is limited and lives on a separate branch. FFmpeg is required for anything that isn’t a WAV.
A CPU-only run at 1.5 times realtime is fine for occasional work. A four minute song takes about six minutes, unattended. For a hundred tracks it’s a weekend, and that’s the point where paying for someone else’s GPU starts making sense.
A licensing catch that matters commercially
Demucs’s code is released under the MIT licence. Its model weights are not. The project states plainly that “the license of the weights is not clear”, citing research-only MUSDB data and internal Meta datasets used in training.
If you’re separating stems for your own practice, remixing at home, or transcription, this is academic. If separated stems are going into something you intend to sell, you’re relying on weights whose licence the authors themselves describe as unclear, and that’s a different risk from a commercial service that has explicitly granted you rights. It’s the same shape of problem we ran into reading what Suno’s terms permit: a tool working is not the same as a tool being licensed for what you’re doing with it.
We’re flagging the ambiguity, not resolving it. Nobody can resolve it from outside.
What we couldn’t price, and why
We won’t publish a price we didn’t read.
Moises doesn’t expose its plan pricing on a crawlable page. Its public pricing page
describes a free tier with “a limited number of uploads each month” and points to a table
inside the logged-in app at studio.moises.ai/billing/pricing, which returned no plan data
to us. It does publish that it separates into 27 stem types.
iZotope RX carries Music Rebalance, the feature usually grouped with these tools, but not in the cheap edition. RX 12 Elements is €109 and does not include it. Music Rebalance starts at RX 12 Standard, €439, and Advanced is €1,499. iZotope also sells Photon, a dedicated stem separation app, at €284.25 reduced from €379. We went through which edition does what in iZotope RX: Which Edition Has Music Rebalance?.
Treat any third-party page quoting current prices for these tools with suspicion unless it says when it checked. The secondary sources we sampled while researching this disagreed with each other, and several were quoting figures over a year old.
We also have no quality comparison between LALAL.AI and Demucs, because LALAL.AI doesn’t publish an SDR figure and we don’t run listening tests. The only head-to-head benchmark we found in this space was published by one of the vendors being compared, which tells you what it’s worth.
Who should pick what
An Nvidia GPU with 8 GB, separating stems regularly. Install UVR. The best published score in this article costs nothing, and your audio never leaves your machine. If you’re working on unreleased material for other people, that last point is an argument on its own.
Stems occasionally, on a laptop. Run Demucs on CPU and go make coffee, or use LALAL.AI’s free 10 minutes. Don’t start a subscription for four tracks a year.
Volume, no patience, no GPU. LALAL.AI Lite at €6.75/month, monthly rather than annual, and accept the Relaxed Queue for anything that isn’t urgent. The unlimited relaxed processing is the part worth paying for.
Deadlines. Pro, monthly, and only if you’ll genuinely use more than 90 Fast minutes a month. The minutes vanish at the end of each period, and the upgrade stops paying for itself the moment you stop consuming it.
Prices here were read on 7 November 2026. We re-check this page roughly every 30 days and change the date at the top when we do. If you spot a figure that’s moved before we do, tell us.
Questions people actually ask
Is there a free stem separation tool that is actually good?
Demucs is free and its documentation reports 9.00 dB SDR on the MUSDB HQ test set for the fine-tuned v4 model, the highest published figure of anything covered here. It needs roughly 3 GB of GPU memory as a minimum, or runs on CPU at about 1.5 times the track's duration.
Does LALAL.AI give a discount for paying annually?
No. As published on 7 November 2026, Lite is EUR 6.75 a month or EUR 81 a year, and Pro is EUR 13.50 a month or EUR 162 a year. Both annual figures are exactly twelve times the monthly price, so paying up front saves nothing.
Do unused LALAL.AI minutes roll over?
No. Fast Queue minutes reset at the start of each billing period and do not carry forward. The Relaxed Queue is unlimited on both paid plans, so only the Fast Queue allowance is metered.
Sources
- iZotope, RX 12 edition comparison, read 9 November 2026
- LALAL.AI, Pricing, read 7 November 2026
- Demucs, source repository and model documentation (Hybrid Transformer Demucs v4), read 7 November 2026
- Ultimate Vocal Remover GUI, source repository and system requirements, read 7 November 2026
- Moises, pricing information page, read 7 November 2026