Mureka V9.5: AI Songs That Finally Hold Together

AI music has a second-verse problem. The intro lands, the first hook is catchy, and then the song forgets what it was about. The chorus comes back a little different. The bridge arrives from another track entirely. The vocal keeps singing, technically on pitch, with no idea that the lyrics just turned sad. You end up with ninety good seconds and a song you can't use.

Mureka V9.5 is Mureka's answer to that, and it's now selectable on each::labs. Mureka describes it as "an improved songwriting brain that brings melodies, harmonies, and arrangements flow more naturally." Less poetically: the update is about the whole song, not the first impression. Structure that holds, arrangements that breathe, vocals that act like they read the lyric sheet, and prompts that actually get followed.

Here's what changed, how it compares with the Mureka models you may already use, and how to get a finished track out of it.

Why Most AI Songs Fall Apart After the Hook

Generating a convincing fifteen seconds of music is a texture problem. Generating three minutes is a planning problem. A real song has an argument: the verse sets something up, the pre-chorus raises the stakes, the chorus pays it off, and the bridge earns the last chorus. Lose the plan and every section sounds fine on its own and wrong in sequence.

Mureka has been attacking that planning problem for a while with MusiCoT, its music chain-of-thought approach, which Mureka says forms "analyzable musical reasoning before audio is generated." V9.5 pushes it further. In its own write-up of the changes, Mureka says MusiCoT now "connects full-song organization with local expression," so the flow between sections and the emotional development across the track stay coherent.

That's the right thing to fix. Nobody listening to a demo cares about one perfect bar. They notice when the song stops making sense.

What's New in Mureka V9.5

Mureka shipped V9.5 in its app on August 13, 2026, and its official breakdown groups the changes into four areas.

Song structure. Sections connect. Verses lead somewhere, choruses return as the same chorus, and the emotional arc builds instead of resetting every sixteen bars.

Arrangement. Density now follows the job of each section. Mureka says the arrangement leaves room for the main melody, the vocal and the emotion, rather than stacking every instrument everywhere. A sparse verse and a full chorus is the oldest trick in songwriting, and it's what makes a chorus feel big.

Vocals. Vocals, lyrics, delivery, emotion and accompaniment work together more consistently. Mureka's launch release talks about smoother phrasing and better interaction between the lead and the supporting parts.

Prompt execution. The genre, mood, vocal direction and arrangement you ask for come through more reliably. If you write "slow, intimate, female vocal, piano and brushed drums," you're more likely to get exactly that and not a polished pop production that loosely agrees with you.

Mureka also published its own evaluation on 100 songs against its previous baseline: a vocal pass rate of 61.0% (up from 53.0%), prompt control at 97.0% (from 92.0%) and genre fully expressed in 95.7% of songs (from 84.9%). These are Mureka's internal numbers, so read them as direction, not a benchmark. The direction is clear though: the biggest jumps are in the two areas people complain about most, vocals and doing what you asked.

Mureka V9.5 vs V9, V8 and O2

If you've built with Mureka before, you'll see a model dropdown with several options. Here's how they relate, from Mureka's own timeline and API docs.

Mureka V8 (January 2026) was the big jump toward real-world use: better musicality, arrangement, vocal expression and audio texture, and vocals that matched prompts better. If you read our Mureka V8 overview, that's the model it describes.

Mureka V9 (March 2026) moved the focus to control: stronger text control within sections, better generation efficiency and better mixing quality.

Mureka V9.5 builds on that with the structure, arrangement, vocal and prompt-following work above. For a new song, it's the one to start with.

Mureka O2 is the O-series reasoning model. The each::labs schema describes it as the strongest for structural coherence and multilingual coverage, but slower, and it doesn't support streaming, a cloned vocal or a melody input. Reach for it when structure matters more than speed and you don't need those inputs.

The rule of thumb: V9.5 by default, O2 for long or structurally demanding pieces, and the older versions only when you need to match a sound you already shipped.

How to Run Mureka V9.5 on each::labs

On each::labs, Mureka V9.5 lives inside Mureka Generate Song. Set the model field to mureka-9.5. The default, auto, picks Mureka's current best regular model, but naming the version keeps your output stable when Mureka updates what auto means.

The inputs are straightforward. Lyrics are required, up to 3,000 characters, in any of ten languages including English, Spanish, French, German, Japanese and Korean. Prompt is the style brief: genre, mood, instruments, vocal delivery, up to 1,024 characters. Gender picks a female or male vocal. n sets how many versions you get back, one to three, default two. That A/B pair is more useful than it sounds; two takes of the same brief tell you quickly which choices were the model's and which were yours.

Three optional controls change what's possible. A reference (a 30-second clip) steers the style. A vocal (a 15 to 30 second clip) clones a voice timbre, and you can make a reusable one with Mureka Vocal Clone. A melody (5 to 60 seconds, MIDI recommended) seeds the composition itself, and when you use it the prompt, reference and vocal are ignored. Turn on streaming if you're building an app and want playback to start before the full song renders.

A clean first request looks like this:

{
  "model": "mureka-9.5",
  "lyrics": "[Verse]\nStreetlights hum on an empty road...\n[Chorus]\nAnd I'm still driving home...",
  "prompt": "indie pop, bittersweet, warm female vocal, clean electric guitar, soft synth pads, slow build into a big final chorus",
  "gender": "female",
  "n": 2
}

Write Lyrics Like a Map, Not a Poem

V9.5's structure gains only help if you give it a structure to follow. The lyrics field supports square-bracket section tags like [Verse], [Pre-Chorus], [Chorus] and [Bridge], and the each::labs schema notes that the model honors them and timestamps them in the response. Use them every time.

Repeat your chorus word for word. If the chorus text changes, you're telling the model it's a different section. Keep verse lines similar in length so the phrasing has a pattern to settle into. Put the emotional turn where you want the music to turn, typically the bridge, and say it in plain words.

No lyrics yet? Mureka Generate Lyrics writes them from a short description of the theme, mood and story, and its output drops straight into Generate Song. Edit the draft before you send it. Thirty seconds of human taste on the chorus line pays off more than any prompt tweak.

Brief the Sound Like a Producer

V9.5 follows prompts better, so a better brief now gets you more. The pattern that works is genre, then mood, then vocal, then instruments, then the shape of the song:

Lo-fi hip hop, late night and reflective, soft male vocal half spoken, dusty piano loop, vinyl crackle, laid back drums, stays minimal until the last chorus adds strings.

Upbeat Latin pop, celebratory, bright female vocal, nylon guitar and congas, handclaps in the chorus, short breakdown before the final chorus.

Cinematic folk, hopeful, intimate male vocal close to the mic, fingerpicked acoustic guitar, cello enters in the second verse, full band only in the bridge.

Notice the last clause in each one. That's arrangement direction, and it's exactly the area V9.5 improved. Telling the model where the song should be sparse and where it should open up is how you get a chorus that hits.

Finish the Track With the Mureka Toolkit

A generated song is a draft. The rest of the Mureka toolkit on each::labs is how you turn it into a deliverable.

Fix one section. Region Edit regenerates a time range with new lyrics and blends it back in, so a weak second verse doesn't cost you the chorus you love.

Make it longer. Extend Song continues a track from a timestamp with new lyrics, for a full-length edit of a short draft.

Split it apart. Stem Song V2 separates a song into 12 stems and extracts each instrument as MIDI, ready for a DAW. Pair it with Generate Track to add a new bass line, string part or vocal that fits the existing song.

Drop the vocal. Generate Instrumental also supports mureka-9.5, for background music under video, podcasts and games.

All of these sit behind the same each::labs API, so chaining them is just more calls, or a visual flow you build once. If you're building a product around this, our guide to building AI song applications walks through the architecture, and the piece on commercial use and licensing for AI music covers the questions your legal team will ask.

Frequently Asked Questions

What is Mureka V9.5 best at?

Whole songs. Mureka focused this release on structure across sections, arrangements that change with each section's role, vocals that match the lyrics' emotion, and following your genre and mood brief. It's the version to use when the track has to work from intro to outro.

How do I use Mureka V9.5 on each::labs?

Open Mureka Generate Song, set the model to mureka-9.5, add lyrics with section tags and a style prompt, and generate. The same version is available in Generate Instrumental for tracks without vocals.

Is Mureka V9.5 better than Mureka V8?

For most new work, yes. V8 brought the big jump in vocals and audio texture; V9 added control within sections and better mixing; V9.5 adds the song-level structure and arrangement work. Keep V8 only if you need to match tracks you already made with it.

Can Mureka V9.5 sing with a cloned voice?

On each::labs you can pass a vocal clip of 15 to 30 seconds to clone a timbre, or create a reusable voice with Mureka Vocal Clone first. Use a clean solo vocal and make sure you have the rights to the voice.