The Cheapest Way to Separate Stems in 2026 (And What You're Actually Paying For)

If you want the cheapest stem separation available right now, here is the honest answer before anything else: for most musicians, the cheapest way to separate stems is free, it runs on your own computer, and there is no cap on how many songs you put through it. You are not missing a discount. Free local splitting is a real option in 2026, and for practice work it is often good enough that paying would be waste.

So the interesting question is not "what is the cheapest AI stem splitter." It is: when free stops being good enough, what exactly are you buying, and what is the smallest amount you can buy? That is what the rest of this is about. I build PracticeSession and Lamina, both of which include free unlimited local separation, so you can weigh my conclusions accordingly. I have tried to write the page I wanted when I was pricing this out.

Four ways to pay for stem separation (and what each one actually costs)

Every tool on the market bills in one of four shapes. Recognising the shape matters more than any individual price, because prices in this category change faster than anyone can write about them.

1. Run it on your own machine: free, unlimited, no account

Separation models can run locally. No upload, no queue, no per-song cost, no monthly cap. The bill is your electricity and some minutes of processing.

Ultimate Vocal Remover (UVR) is the well-known free option. Open source, Windows, Mac and Linux, and it can run Demucs and the newer transformer models that top the quality tables. It is genuinely free with no asterisk. The cost is setup: you install it, download model files, pick a model from a list that assumes you know what the acronyms mean, and it strongly prefers an NVIDIA GPU. On CPU alone it works but it is slow. If you enjoy that kind of thing, UVR is the best value in audio software, full stop.

PracticeSession and Lamina both include local separation using Spleeter models, free and unlimited, with no per-song charge and nothing to configure. Lamina is a free standalone splitter, so if all you want is stems on disk, that route costs nothing at all. Spleeter is an older engine than what UVR can run, and it shows on dense mixes, but for isolating a bass line to transcribe or muting a guitar to play along, it does the job more often than people expect.

The honest catch with local separation: older models, and processing time proportional to your hardware. Quality is where your money goes in this market, and free local is the bottom of the quality ladder unless you are running UVR with good models on a real GPU.

2. Free web tiers: free until a cap you did not read

Nearly every cloud service has a free tier, and the free tiers have quietly become the most competitive part of the market. What varies is the cap, and the cap is the part worth checking before you invest any time.

Caps come in five flavours, and services mix them: uploads per month, maximum track length, output format (MP3 free, WAV paid), file expiry, and stem count. Two current examples, both read on the vendor's own pages in August 2026:

• Moises allows five file uploads per month on its free plan, with the counter resetting each monthly cycle. Five songs a month is a real constraint if you are learning a set.
• Fadr takes the opposite approach: its free tier is uncapped for basic stems with MP3 download, and Fadr Plus at $10 per month adds finer stem types and lossless WAV.

A wave of no-signup web splitters has also appeared in the last year. They are fine for a one-off. Read what happens to your file, check whether the output is MP3 only, and do not build a workflow on a site that might not exist next spring.

3. Per-minute credits: you pay for audio, not for songs

The serious cloud engines mostly meter by minutes of audio processed. You buy a block of minutes, they come off the block as you use them, and better packs cost less per minute. LALAL.AI popularised this shape; PracticeSession and Lamina use it too for their higher-quality cloud separation.

Per-minute billing has one property nobody advertises, and it is the single most useful thing in this article: if you pay by the minute, feeding the machine less audio costs less money. More on that below, because it is where most of the savings actually are.

For reference, PracticeSession and Lamina share the same credit packs: 20 minutes for $3.99, 60 minutes for $7.99, or 150 minutes for $15.99, which works out between roughly $0.11 and $0.20 per minute of audio depending on pack size. Credits are shared across both apps.

4. Subscriptions: cheapest per song, worst if you are seasonal

Monthly plans are the cheapest per song for anyone processing constantly, and the worst possible deal for how most musicians actually work. Learning happens in bursts: you take a gig, you learn eleven songs in three weeks, and then you do not touch a splitter for two months. A subscription bills all five of those months.

Check two things before subscribing. First, whether the monthly allowance rolls over, because it usually does not, so a light month is money burned. Second, whether the stems remain accessible if you cancel, or whether they live in the service.

Four ways to spend less without accepting worse stems

Assume you have decided free local separation is not clean enough for the song in front of you. These four moves cut the bill without touching quality, and the last two are the ones almost nobody uses.

Start local, escalate only when it fails

Run the free local split first. Listen to the part you actually care about. Most of the time it is fine, and you have spent nothing. When it is not fine, and you will know within about ten seconds, spend on the cloud engine for that one song. Treating paid separation as an exception rather than a default is the largest single saving available, and it costs you one listen.

Separate the region, not the whole song

You are learning a sixteen bar solo. The song is four minutes long. Under per-minute billing, separating the whole song costs you four minutes of credit to get thirty seconds you will use. Separate only the region and you spend roughly an eighth of that.

This is the saving people leave on the table most often, because uploading the whole file is the path of least resistance on a web tool. In PracticeSession you can set a region and separate just that region, which is a habit worth building the first week you own it.

Isolate one instrument instead of buying a full split

Most of the time you do not want stems, plural. You want the bass, so you can hear what it is doing. A full multi-stem split gives you vocals, drums, bass and the rest, charges you a multiple of the track length for the privilege, and you listen to one of them.

Single-instrument isolation solves this. In PracticeSession and Lamina, isolating one instrument costs 1× the audio duration: a three minute passage costs three minutes of credit. A full multi-stem split of the same audio costs a multiple of that. If you only need the piano, pay for the piano.

Keep the stems, so you never pay for the same song twice

The most expensive stem is the one you separate twice. It happens constantly with web tools: the tab is closed, the file expired, the download folder has four files called bass.wav, and it is easier to re-run the song than to find them. You have now paid twice for identical audio.

Two defences. Check whether the tool writes stems to your own disk in a location you control, rather than keeping them inside a service. And prefer a tool that keeps the stems attached to the song rather than loose in a folder. In PracticeSession, stems belong to the song in your library, so reopening the song reopens its stems, still lined up with your loops and your notes. There is nothing to re-find and nothing to re-pay for.

The cost nobody puts on the pricing page

Every price in this market is quoted per split. That is not the number that matters. The number that matters is the cost of a stem you can use, and between those two numbers sits a chunk of unpriced work.

A stem in a downloads folder is not finished. To practise with it you still have to import it somewhere, line it up against the original, find your section again, set your loop again, and get the balance right between the part you are learning and the part you are playing over. That is five to ten minutes per song, every song, and it recurs every time you come back to the material.

Run the arithmetic on a modest set. Eleven songs, a few dollars of separation cost, and something close to two hours of shuffling files. If your time is worth anything at all, the file shuffling is the larger expense, and it is the one no comparison table has ever counted. The cheapest tool by sticker price can easily be the most expensive tool by the hour.

Which is the actual argument for separating inside a practice app rather than beside one. When the stems arrive attached to the song, with a mixer, your saved loops, your slow-down settings and your notes already on it, the post-processing cost is zero because there is no post-processing.

So which one should you use?

If what you want is a straight quality ranking rather than a cost argument, that is a different article and I wrote it separately: Best AI Stem Separation Apps Compared puts Moises, LALAL.AI, Gaudio, Fadr and the rest side by side on output quality and features.

Prices and free-tier caps in this article were checked against vendors' own pages in August 2026, and I have deliberately printed very few of them: several companies in this category have moved pricing behind a login, and a figure I cannot re-verify is a figure I should not quote. If something here has gone out of date, tell me and I will fix it.

Frequently Asked Questions

What is the cheapest way to separate stems from a song?

Running separation locally on your own computer, which is free and has no monthly cap. Ultimate Vocal Remover is the free open-source route and reaches the highest quality if you have an NVIDIA GPU and are willing to configure it. PracticeSession and Lamina both include free unlimited local separation with no setup. Paid cloud separation is worth buying only for the specific songs where the free result is not clean enough.

Is there a free stem splitter with no limits?

Yes, if you run it locally: local separation has no upload cap, no track-length cap and no monthly quota, because nothing is being processed on someone else's server. Free web tiers almost always cap something. Common limits are uploads per month, maximum track length, MP3-only output, and files that expire. Moises allows five uploads per month on its free plan; Fadr's free tier is uncapped for basic stems with MP3 download. Both figures were current in August 2026.

Is stem separation cheaper per song or per minute?

Most serious cloud engines charge per minute of audio, not per song, which means a short song costs less than a long one and, more usefully, that separating only the section you care about costs a fraction of separating the whole track. If you are paying per minute, separating a sixteen bar passage instead of a four minute song is roughly an eighth of the cost for exactly the material you will use.

Should I pay for a stem separation subscription?

Only if you process constantly all year. Most musicians work in bursts: a run of songs before a gig, then nothing for weeks. A subscription bills the quiet months too, and monthly allowances usually do not roll over. Non-expiring per-minute credits fit that pattern better. Check as well whether your stems stay accessible after you cancel.

Do I need to pay for a full multi-stem split?

Usually not. Most of the time you want one instrument, not four files. Single-instrument isolation in PracticeSession and Lamina costs 1× the audio duration, so a three minute passage costs three minutes of credit, while a full multi-stem split of the same audio costs a multiple of that. If you only need the bass line, isolate the bass line.

Is free stem separation good enough for practice?

More often than people assume. Practice does not need a broadcast-clean stem; it needs to hear a part well enough to copy it, or to get your own instrument out of the way so you can play along. Free local separation handles that on most material. Dense mixes, heavy production and close-harmony vocals are where it audibly struggles, and those are exactly the songs worth spending cloud credits on.