Correction, 2026-09-30: MiniMax Music 3 was listed at a flat 17 credits per generation; it bills per second (0.27 credits), so 17 is a 60-second track and a 3-minute track is 49.
A song is music plus vocals plus lyrics, and that combination is where AI audio got genuinely surprising in 2026. Free tiers exist almost everywhere. The thing that should decide your choice is not output quality, which is broadly similar, but what licence the free tier grants you, because a free-tier track used in a monetised video is the most common way people create a rights problem for themselves.
TL;DR
- Free tiers: usually a monthly generation allowance, non-commercial licence, and sometimes no download
- The trap: free almost always means personal use only. Commercial use needs the paid tier, and that includes a monetised YouTube video
- On 8frame: MiniMax Music generates music with lyrics from 5 credits (v2) or 17 (Music 3). ElevenLabs Music is 81 credits per minute. Lyria 2 is 13 credits per 30 seconds
- Vocals are the weak link. Instrumentals hold up better than sung lines, especially in the top of the register
- Never clone a real artist's voice. Not a grey area
What free tiers actually grant
The generation allowance is rarely the constraint. The licence is. Typical shape:
- Free: personal, non-commercial use. Your own project, not a client's, not a monetised channel, not an ad.
- Paid: commercial use, sometimes with a distinction between using it in content and distributing it as music.
- Ownership language varies a lot. Some services grant you ownership, others a licence they can change.
Read the current terms on the service itself rather than any summary. This is the single most consequential thing in the category and it changes.
What it costs on 8frame
Canon prices, a credit is $0.01 at pack rate:
| Route | Credits | What it is |
|---|---|---|
| MiniMax Music v2 | 5 | Music with lyrics, the cheapest route |
| MiniMax Music 3 | 17 for 60 s, per second | Newer tier, same family |
| Lyria 2 | 13 per 30 seconds | Google's music model |
| ElevenLabs Music | 81 per minute | The premium option |
| ElevenLabs SFX | 1 to 6 by length | One-off sound, not music |
| ElevenLabs TTS | 13 per 1,000 characters | Spoken, not sung |
A 3-minute track: ElevenLabs Music at 81 per minute is 243 credits, about $2.43. On Lyria 2 at 13 per 30 seconds it is 78 credits. MiniMax Music 3 at 49 for three minutes, billed per second, is cheaper still, and the reason to pay more is consistency across a longer piece.
Where vocals still fail
The top of the register. Sustained high notes are where synthetic vocals thin out and the artefacts become audible.
Consonants at speed. Fast lyrical passages blur. Write fewer syllables per bar than you would for a human.
Emotional turns. A vocal that needs to break, hold back, or land a single word carries the song, and that is exactly what does not arrive reliably.
Long-form consistency. A voice that sounds right for 30 seconds can drift over three minutes. Generating in sections and assembling is more reliable than asking for the whole thing.
What works well: instrumental beds, hooks, stylistic pastiche, demo versions of an idea you will later record properly, and anything where the track is under dialogue rather than in front of it.
The rule with no exceptions
Do not generate a specific artist's voice or a specific artist's song. Prompting toward a named living performer's vocal identity is both a legal exposure and, on most services, a terms violation. The same applies to a recognisable song's melody.
Stylistic reference is different from identity: "1970s soul, warm tape saturation, female lead" is a description. Naming a person is not.
The honest place this fits
Good fit: background beds under video, hooks for social content, sound for a product demo, a demo of a songwriting idea, and any case where licensing a library track would cost more than the value of the track.
Bad fit: the song being the product. If people are meant to listen to it as music and care who made it, the thing they care about is the person, which is the part that cannot be generated.
The comparison nobody makes: a library subscription gives you human-performed tracks with clean commercial licences for roughly the price of a plan here. The reason to generate instead is specificity, that you want this exact feel at this exact length, not cost.
FAQ
Can I monetise a video with a free-tier AI song? Usually not. Free tiers are typically personal use. Check the service's current terms.
Can AI write lyrics too? Yes, and it is the weakest part of the output. Write them yourself; it costs nothing and it is the only part where you have an advantage.
Which is cheapest on 8frame? MiniMax Music v2 at 5 credits per generation. ElevenLabs Music at 81 per minute is the one to pay for when the track is exposed.
Sources
- MiniMax Music 3 billing: 8frame tool registry credit function, read 2026-09-30
Generate the bed, write the words yourself. The 8frame canvas is free and unlimited; generation is paid from $19/month.