Background Music Models Ranked (2026): Vocals, Instrumental and Prices

A background music bed sets the pace of a film, and a vocal track can carry it. This ranking starts from the Artificial Analysis music leaderboards for instrumental and vocals — published preference rankings from blind comparisons — and filters them to the models North Noir can provide through ElevenLabs and fal, with provider prices and our own test generations alongside.

By Syver-August, Developer, North Noir · October 17, 2026 · 8 min read

The leaderboards, and what they do and don’t say

Artificial Analysis runs published preference rankings for music: blind comparisons by validated private evaluators, reported as rank, Elo score and a 95% interval, on separate vocals and instrumental boards (captured 17 October 2026). Two caveats travel with every number below: the intervals overlap heavily, so a one-place gap is not proof of a quality gap; and a preference ranking of full songs is not a ranking of 18-second beds.

We filtered both boards to the models we can actually provide through ElevenLabs and fal, and kept the leaderboard ranks so the order stays auditable. The “Price” column is the provider’s published rate — the leaderboard itself quotes none — and the verdict column is our visual/metric review of what the route delivered in our own runs, not a listening result.

Instrumental — AA-Music-Instrumental v1.1 ranks 3–18, filtered to providable models. Scores and ranks captured 2026-10-17; prices from provider pages the same day.
Rank	Model	AA score	Runs on	Price	Our test verdict
#3	Mureka V9.5	Elo 1084	fal	$0.225/request instrumental	Highest instrumental score we can run; our runs returned full songs (166–179s), not short beds
#4	Mureka V9	Elo 1083	fal	$0.0675/request	Full songs again (173–199s); same song/instrumental routes with `model: mureka-9`
#5	Lyria 3 Pro	Elo 1080	fal	$0.08/generation	One of our two prompts was rejected by provider moderation; the other returned 48.3s
#6	Lyria 3.5	Elo 1078	fal	$0.10/generation	No duration field: our 18s target hint produced 62–89s outputs
#7	Eleven Music	Elo 1021	ElevenLabs direct + fal	$0.15/minute direct	Top of our timing-fit composite (67.2): 18.02s vs an 18s target
#9	Eleven Music v2	Elo 1011	ElevenLabs direct + fal	$0.15/minute direct	18.05s vs 18s; composite 65.0
#10	MiniMax Music 2.5+	Elo 1007	fal	$0.15/generation	No duration control in the schema: 92–120s outputs in our runs
#11	Eleven Music v2.5	Elo 1001	ElevenLabs direct + fal	$0.15/minute direct	18.05s vs 18s; composite 62.9
#12	MiniMax Music 2.6	Elo 1001	fal	$0.15/generation	Long arrangements again (93–279s) — no numeric duration input
#13	MiniMax Music 3.0	Elo 1000	fal	$0.002/second	18.01s when a duration bound is set; the cheapest timed bed we tested
#14	Stable Audio 3 Medium	Elo 996	fal	$0.0376/generation	18.00s exactly; our strings cue lost its main body near the end of the clip
#16	Stable Audio 2.5	Elo 984	fal	$0.20/generation	18.00s exactly; a visible continuity warning on the strings cue in our spectrogram
#17	Stable Audio 3 Small	Elo 946	fal	$0.0217/generation	18.00s exactly and cheap per clip; lowest published score of the providable set
#18	Lyria 2	Elo 926	fal	$0.10/30 seconds	Fixed 32.77s outputs — a 30-second route, not a short-cue one

Vocals — AA-Music-Vocal v1.1 ranks 3–12, filtered to providable models. Scores and ranks captured 2026-10-17; prices from provider pages the same day.
Rank	Model	AA score	Runs on	Price	Our test verdict
#3	Mureka V9.5	Elo 1107	fal	$0.225/request lyrics-to-song (prompt-only $0.75)	The best published vocal score we can run; our two runs returned full songs (182–202s)
#4	Mureka V9	Elo 1057	fal	$0.0675/request (prompt-only $0.45)	Full songs in our runs (153–158s); identical routes with `model: mureka-9`
#5	Lyria 3.5	Elo 1052	fal	$0.10/generation	88–179s outputs from an 18s target hint; no duration parameter in the schema
#6	Lyria 3 Pro	Elo 1051	fal	$0.08/generation	64–90s outputs; two of two sea-chorus runs delivered
#7	Eleven Music v2.5	Elo 1012	ElevenLabs direct + fal	$0.15/minute direct	18.05s vs an 18s target — the exact-length option with the highest vocal score
#8	MiniMax Music 3.0	Elo 1000	fal	$0.002/second	18.01s with a duration bound; lyrics are required even for instrumental tags
#9	Eleven Music v2	Elo 999	ElevenLabs direct + fal	$0.15/minute direct	18.05s vs 18s
#10	Eleven Music	Elo 999	ElevenLabs direct + fal	$0.15/minute direct	18.02s vs 18s
#11	MiniMax Music 2.6	Elo 998	fal	$0.15/generation	39–54s outputs in our runs
#12	MiniMax Music 2.5+	Elo 985	fal	$0.15/generation	44–190s outputs in our runs

Our test generations

We asked each route for the same small set: two instrumental prompts — “warm reflective piano, slow, hopeful” and “tense low strings for a mystery scene” — and two independent runs of one original vocal request, “gentle folk chorus singing about the morning sea”, with the same four-line chorus where the route requires supplied lyrics. The target was 18 seconds where the API accepts a length at all; where it does not, we kept the full uncropped song instead of cropping it to disguise that. We reviewed spectrograms and measurements — we did not listen, verify the sung words, or test mixing quality.

Prompt: “tense low strings for a mystery scene” (instrumental). 18.05s delivered against an 18s target — the length discipline that wins this route its top timing composite.

Spectrogram of the Eleven Music v2.5 instrumental strings clip showing repeated attacks over a sustained harmonic body across eighteen seconds — The instrumental clip’s spectrogram: regular attacks over a sustained body, energy declining smoothly after ~14s.

Prompt: “gentle folk chorus singing about the morning sea”. Full song (~202s) from a route with no duration control — highest published vocal preference score we can run (AA #3, Elo 1107).

The two sample families show the real trade-off. Mureka V9.5 tops the published preference ranking among the models we can provide and returns whole songs. Eleven Music v2.5 sits lower on the board but gave us exact 18-second beds on every run — which is what a timeline wants. Neither fact cancels the other.

Prices, and why the route matters

Eleven Music is metered differently on each route: the direct API publishes $0.15/minute, while the fal pages for the same model family publish $0.60 per output minute, rounded up to a whole minute — so an 18-second fal generation is quoted as $0.60. Copying the direct price onto a fal call is the easiest way to under-estimate a budget by 4×. Mureka is metered per request instead: $0.225 for lyrics-to-song or instrumental on V9.5 ($0.0675 on V9), and prompt-only songs cost more — $0.75 and $0.45 respectively.

The cheap end is per-second: MiniMax Music 3.0 at $0.002/second of output, so a bed simply meters at that rate times its length. Stable Audio routes charge per generation ($0.02–$0.20 in the tables). All figures are published rates checked 17 October 2026, not invoices.

How North Noir music beds work

In the Studio, the music lane has a Generate with ElevenLabs block: describe the bed (“dark sparse piano, slow, cinematic, no vocals”), pick a length, and it lands as an instrumental loop under the narration lane on the editable timeline — charged by length in tokens. Music is a separate lane from narration and sound effects, so re-timing or muting it never touches the drawn frames.

Outside the Studio, POST /v1/audio/music generates a bed from a prompt, and the pricing page shows the token cost. On the free plan the music provider runs on your own API key — an ElevenLabs key covers the Studio’s music lane and a fal key opens the other routes in these tables.

Exact short beds

Eleven Music (direct, $0.15/minute): music_length_ms and force_instrumental made 18s beds exactly
MiniMax Music 3.0 ($0.002/second) with an explicit duration bound
Best fit under narration, where the bed must end with the section

Vocal tracks and full songs

Mureka V9.5: the highest published vocal score we can provide (AA #3, Elo 1107) at $0.225/request
Mureka V9 at $0.0675/request when you want the same routes cheaper
Both return whole songs — plan to cut them on the timeline

What we left out

Five leaderboard names are not in the tables because we cannot provide them through ElevenLabs, fal or OpenAI: Suno v6 and Suno v6-mini (the top two of both boards), Stable Audio 3 Large (no Large endpoint in the catalog; the Small and Medium routes are not substitutes for its score), Melodia v3 and MusicGen. OpenAI’s audio API covers speech and transcription, not music generation, so nothing from it appears here either.

For the cue half of a film’s sound — one-shots and ambience — read the companion piece, Sound Effects Models Ranked.

Key takeaways

The Artificial Analysis music boards are published preference rankings from blind comparisons — but they publish no prices, and their confidence intervals overlap, so ranks are not proof of a quality gap.
Among the models we can provide, Mureka V9.5 has the highest published scores on both boards (vocals #3/1107, instrumental #3/1084) and returns full songs at $0.225/request.
Eleven Music (direct, $0.15/minute) delivered exact 18-second beds in our tests and topped our timing composite — the practical pick under narration.
Route changes price: fal rounds Eleven Music up to a whole output minute ($0.60), MiniMax Music 3.0 meters at $0.002/second with a duration bound.

Frequently asked questions

What is the best AI music model we can actually use?

By published preference score, Mureka V9.5 is the strongest we can provide — Artificial Analysis ranks it #3 on vocals (Elo 1107) and #3 on instrumental (Elo 1084). By fit for an exact-length bed, Eleven Music direct won our tests at 18.02–18.05s against an 18s target. The leaderboards’ confidence intervals overlap, so treat close ranks as ties.

How much does AI music generation cost?

Published rates checked 17 October 2026: Eleven Music direct is $0.15/minute, Mureka is $0.225 per lyrics-to-song or instrumental request on V9.5 ($0.0675 on V9, with prompt-only songs costing more), MiniMax Music 3.0 is $0.002/second, and the Stable Audio routes charge $0.02–$0.20 per generation. The fal pages for Eleven Music round up to a whole output minute at $0.60.

What is the Artificial Analysis music leaderboard?

A published preference ranking built from blind comparisons by validated private evaluators, reported as rank, Elo score and a 95% confidence interval on separate vocals and instrumental boards. It ranks models by how often evaluators preferred their output; it publishes no prices, and overlapping intervals mean small rank differences are not statistically meaningful.

Can I get an exact-length background music bed?

Only from routes that expose a length: Eleven Music takes music_length_ms (our runs landed within a tenth of a second of 18s) and MiniMax Music 3.0 accepts a duration bound. Mureka, Lyria and several MiniMax routes have no duration parameter and returned full songs of 39 to 279 seconds in our tests, which you then cut on the timeline.

How does North Noir generate background music?

The Studio’s music lane has a Generate with ElevenLabs block that makes an instrumental bed of a length you pick, charged by length in tokens and placed under the narration lane on the editable timeline. The same capability is exposed as POST /v1/audio/music in the API.

Open the Studio