All posts
Music Generation API

Mureka V8 vs V9.5: Which Model to Pick for Your Song

Mureka V8, V9, V9.5, O2, 7.6 or auto? A side-by-side verdict built from Mureka's own release notes, plus a two-render A/B test, a tagged lyrics template and eight style prompts to try.

By Eftal Yurtseven · Nov 20, 2026 · 9 min read

Mureka V8 vs V9.5: Which Model to Pick for Your Song

Every Mureka request carries one small field that quietly decides whether your chorus lands: model. Most people leave it on auto, some hard-code whatever was newest the week they integrated, and almost nobody goes back to check. If you searched for Mureka V8, you're probably at that exact fork: V8 is the model you know, V9.5 is the one everybody's posting about, and the dropdown in Mureka Generate Song on each::labs offers seven values. This is the side-by-side verdict, built only from what Mureka has actually published, plus a way to settle the rest with your own ears.

No invented scores here. Mureka's notes on these models are short, and where they stop, this guide switches from "which is better" to "how to find out for your song in two renders".

What Mureka V8 Actually Changed

Mureka V8 arrived at the end of January 2026, and its pitch was ambitious: move AI music from "generatable" to "publishable". Launch coverage on AIbase listed the upgrades in four places. Melodies became catchier, with complete section structure and a clearer split between main and secondary parts. Vocal tone and singing style were said to follow the prompt more closely, with more confident, smoother performances. Arrangements got fuller instrumentation and a more natural emotional build. And the mix itself was described as better balanced, with clearer vocals and a more layered sound field.

Under the hood, Mureka credited its MusiCoT system, which models musical structure and expressive intent before audio is rendered. That's why V8 felt like a jump rather than a patch: the song was being planned, not just continued. Then on March 2, 2026, Mureka's API changelog logged an enhanced version of mureka-8, so the V8 you call today isn't the V8 from launch week.

On each::labs, the request schema describes mureka-8 as the flagship regular model with the best all-round quality. Read that as a reputation, not a ranking. All-round means it rarely surprises you in either direction, which is exactly what you want from a default you can build a product on.

A newer voice is not automatically a better one. It is a different one.
A newer voice is not automatically a better one. It is a different one.

Mureka V9 and V9.5: What Mureka Says Moved

V9 came next. The each::labs schema describes mureka-9 as the newest flagship with further improved arrangement and vocals, and Mureka's changelog records an enhanced mureka-9 on April 9, 2026. That's the whole official story for V9.

V9.5 has more on the record. Mureka's API changelog logged the enhanced mureka-9.5 on August 28, 2026, and the company's September 1 press release framed the release around three areas: arrangement logic, vocal performance and musical consistency. In Mureka's words, sections, instruments and vocals develop with "a clearer sense of structure and musical direction", and vocals get "smoother phrasing, stronger emotional delivery, and better interaction between lead vocals and supporting elements".

Mureka also published internal numbers: lead vocals judged convincing in 61.0% of cases, prompts followed in 97.0% of generations, and a 95.7% match rate with requested genres. Those are the company's own tests and they don't include a V8 baseline, so treat them as a claim about V9.5, not as a margin over anything. The stronger signal is behavioral: on November 12, 2026, Mureka made mureka-9.5 the default for its Remix API and told developers to set "model": "mureka-8" if they wanted to stay on V8. When a model maker moves its own default, that's a vote.

If you want the long read on what V9.5 does to song structure, I wrote it up in Mureka V9.5: AI Songs That Finally Hold Together.

Every stop changes the sound. Pull one on purpose.
Every stop changes the sound. Pull one on purpose.

Mureka V8 vs V9.5: Every Model Value, Decoded

Here's the full menu as the each::labs schema defines it, in plain words.

auto is the current best regular model and the recommended default. It moves when Mureka moves, which is great for a hobby project and risky for a product that needs last month's song to sound like this month's. mureka-7.6 is the fast, efficient regular tier, released with O2 in December 2025. (Send mureka-7.5 and you'll be upgraded to 7.6 transparently, since Mureka deprecated it.) mureka-8 is the all-rounder. mureka-9 pushes arrangement and vocals further. mureka-9.5 is the release Mureka is putting forward for natural phrasing and section-to-section consistency.

Then there's the odd one out. mureka-o2 is an O-series reasoning model, descended from mureka-o1, which Mureka called its first "thinking" music model. According to the schema, O2 has the strongest structural coherence and multilingual coverage of the set, but it's slower and doesn't support streaming. That trade is the whole decision with O2: you give up instant playback for a model that plans harder.

Everything else in the request stays the same whichever model you pick: lyrics up to 3,000 characters in 10 languages (English, Chinese, Japanese, Korean, Portuguese, Spanish, German, Italian, French, Russian), a style prompt up to 1,024 characters, n from 1 to 3 variants, an optional gender of female or male, and three upload hooks: reference_id for a 30-second style reference, vocal_id for a 15 to 30 second voice clip, and melody_id for a 5 to 60 second melody seed (MIDI recommended). That shared surface is what makes a fair A/B possible.

When to Pick Which: The Verdict

Pick mureka-9.5 when the vocal is the product. Ballads, singer-songwriter cuts, anything where a listener judges the song by whether the voice sounds like it means it. That's the axis Mureka says it improved most, and it's where you'd hear a difference first.

Pick mureka-8 when you need a stable, proven baseline. If your app already ships V8 songs, your users have an ear for that sound, and switching mid-campaign changes the brand of your audio. V8 is also the sensible control in every comparison you run, because it's the one most teams have history with.

Pick mureka-9 when arrangement density matters more than vocal nuance, say for a busy pop or electronic track where the instrumentation carries the energy. Treat it as the middle step between the two above.

Pick mureka-o2 for long, multi-section songs and anything outside English. Structural coherence and multilingual coverage are exactly what it's credited with, so a five-part Spanish song with a bridge and a final double chorus is its natural job. Just don't put it behind a "play now" button.

Pick mureka-7.6 for drafts, previews and volume: testing lyric drafts, checking whether a hook scans, generating many options before committing the winners to a flagship. And leave auto for exploration. My position: in production, always pin a version. A model field that changes without a deploy is a bug report waiting to happen.

Two takes per model. Then trust your ears, not the hype.
Two takes per model. Then trust your ears, not the hype.

Settle It With Your Own A/B, Using n=2

Here's the trick most people miss. The schema defaults n to 2 precisely so every request gives you a pair. Use that. Run the same lyrics and the same prompt through mureka-8 with n: 2, then through mureka-9.5 with n: 2. You now have four songs: two takes per model, so you can tell the model's character apart from one lucky seed.

Keep everything else frozen. Same gender, same reference_id if you use one, streaming off for both so the comparison is about the final render. Then listen for three things in order: does the chorus arrive where your [Chorus] tag says it should, does the vocal phrasing breathe or rush, and does the last section feel like an ending or just a stop. Write down which model won each question. After three songs in your genre, you'll have a better answer than any review can give you, because it's an answer about your music.

One more lever: pair a vocal_id with your prompt. Mureka has allowed vocal_id and prompt together since January 2026, so you can hold the voice timbre constant across both models and judge only the writing and the arrangement.

Write the Lyrics Like a Map

Whatever model you pick, structure tags do more work than adjectives. The schema says the model honors square-bracket tags like [Verse], [Chorus] and [Bridge] and timestamps them in the response, which also makes your A/B easier to check. Mureka's changelog notes the label for a breakdown section is [Break]. Here's a shape that gives every model a clear plan:

[Intro]

[Verse]
Streetlights hum a song I almost know
Your jacket on the chair, the radio low
I keep the window cracked to hear the rain
Like every drop is saying back your name

[Pre-Chorus]
And I don't want the morning yet

[Chorus]
Stay till the city lets go of the night
Stay till the last train pulls out of sight
I don't need forever, I just need tonight
Stay till the city lets go

[Verse]
Coffee rings and postcards on the floor
We swore we'd see the coast, we never saw
But here inside this small and borrowed room
The whole wide world is waiting on the moon

[Chorus]
Stay till the city lets go of the night
Stay till the last train pulls out of sight
I don't need forever, I just need tonight
Stay till the city lets go

[Bridge]
If the sun comes up, let it find us here
Half asleep and nothing left to fear

[Break]

[Chorus]
Stay till the city lets go of the night
Stay till the city lets go

[Outro]

Short lines, a repeated chorus with identical wording, and a stripped-back final chorus. That last move is a good stress test: a model with strong structure should drop the energy where you trimmed the words.

A good prompt is a pinned cylinder. It plays what you set.
A good prompt is a pinned cylinder. It plays what you set.

Eight Copy-Ready Style Prompts

These go in the prompt field. Each names genre, tempo feel, instrumentation and vocal delivery, because that's what the parameter is built to steer. Run any of them on two models and you've got your comparison. For a bigger library, there are twenty more in Mureka Prompts: 20 Copy-Ready Song Prompts That Work.

indie pop, mid-tempo, warm analog synths, clean electric guitar, soft live drums, intimate female vocal, close-mic, bittersweet, big singalong final chorus
soul ballad, slow 6/8 feel, upright piano, brushed drums, hammond organ swells, passionate male vocal with gospel runs, restrained verses, soaring chorus
synthwave, driving, 80s gated drums, arpeggiated bass, bright lead synth, cool detached male vocal, night drive, neon, steady build into the bridge
acoustic folk, fingerpicked steel-string guitar, light banjo, soft harmonica, gentle female vocal with breathy harmonies, campfire warmth, sparse arrangement
latin pop, reggaeton groove, nylon guitar hook, punchy percussion, confident female vocal in Spanish, summer energy, catchy chant in the chorus
cinematic rock anthem, slow build, distorted guitars, orchestral strings, huge tom fills, powerful male vocal, quiet first verse, explosive final chorus
lo-fi r&b, laid back, dusty keys, soft vinyl crackle, deep sub bass, smooth male vocal, conversational phrasing, late night mood
j-pop, upbeat, bright piano and strings, bouncy bass, energetic female vocal in Japanese, anime opening feel, key lift into the last chorus

The Spanish and Japanese prompts are the ones to send to mureka-o2 as well, since multilingual coverage is its stated strength. The soul ballad and the folk track are where V9.5's vocal claims either show up or don't.

Beyond Generate Song: Finish the Track

Picking the model is step one. The rest of the Mureka family on each::labs covers what happens after the render. Extend Song grows a track you already like instead of regenerating it. Stem Song splits a finished song into parts for mixing or video edits. Generate Instrumental takes the same model choices when you need a bed with no vocal, and Generate Lyrics drafts the tagged structure for you when the blank page is the bottleneck.

Chain them and you've got a pipeline: lyrics, two songs per model, pick the winner, extend it, split the stems. You can wire that up as a reusable flow in each::labs AI Flows, or browse every model Mureka has on the platform from the Mureka provider page. For the dates behind every model in this post, Mureka keeps an official API changelog.

Frequently Asked Questions

Is Mureka V8 still worth using now that V9.5 is out?

Yes, as a baseline. Mureka still supports mureka-8 explicitly, even telling Remix developers how to stay on it, and the each::labs schema calls it the best all-round regular model. If your product already sounds like V8, switching is a creative decision, not a mandatory upgrade.

What is the main difference between Mureka V8 and V9.5?

Focus. V8's launch was about making songs publishable across melody, vocals, arrangement and mix. V9.5's release targets arrangement logic, vocal performance and consistency, with smoother phrasing and better interplay between lead and supporting vocals. Mureka hasn't published a direct V8 versus V9.5 benchmark, so run the n=2 test on your own lyrics.

Which Mureka model should I use for non-English songs?

Start with mureka-o2. The schema describes it as having the strongest multilingual coverage and structural coherence, at the expense of speed and streaming. All models accept lyrics in the same 10 languages, so A/B it against V9.5 for your language before you commit.

Does auto pick V9.5 for me?

The schema only promises that auto uses the current best regular model, without naming it, and that can change over time. If you need repeatable results, set the model explicitly.

Start with what you need.

Scale when you’re ready.