When AI music companies launch a new model, the headline is almost always the same old song: it'll bring you better vocals, better quality, and better prompt following. But most modern AI music models already sound so good that small improvements can be hard to notice — unless, that is, you compare them side by side. And that's exactly what I've been doing.
I've been using Mureka V9 for the past few months, and it was already a capable model. It generates complete songs, follows prompts reliably, and it gives creators a surprising amount of control over lyrics and musical direction.
V9.5 delivers all of that and more. What stood out to me is hard to explain but something you're definitely going to notice: V9.5 sounds significantly more human, creating songs that feel complete and fully formed. It keeps surprising me with its ability to turn even the wildest ideas into music that feels fresh while still sounding professional.
I decided to create some tests to demonstrate the differences between Mureka V9 and V9.5, and I've included the audio files so you can hear the results.
Test 1: Does V9.5 stay closer to the prompt?
For my first test I kept the prompt simple. I asked for a restrained song featuring soft piano, atmospheric synths, gradual strings and an intimate male vocal that builds slowly from quiet storytelling to an emotional release.
That kind of request looks easy on paper, but it often exposes how some AI models interpret prompts to create music. Many of them understand the instructions but then tend to override them, for example by going for a big chorus and drifting towards overly familiar songwriting patterns.
I used identical prompts for V9 and V9.5, and the results were really interesting.
More natural songwriting
While V9's song was polished, it slowly leaned towards a more conventional pop structure. As the song progressed the arrangement became busier, the vocal became more dramatic, and the emotional arc built to a large climax that the prompt hadn't asked for.
V.95 was significantly more restrained and stuck much better to the brief. The arrangement kept plenty of room for the music to breathe, the emotional progression felt more natural, and the song resisted the temptation to add unnecessary intensity. Instead of forcing the song to follow a familiar, predictable structure, it kept the mood established by the prompt.
More lifelike vocals and mixing
When I listened to both versions back-to-back, there was a noticeable difference in the vocals and in the overall mix too. While both versions' vocals were technically accurate, the V9.5 version had more feel and more life to it: subtle details in the phrasing, timing and dynamics made the vocal more expressive without sounding exaggerated.
The audio mix was a more involving listen too. The various instruments occupied the space more naturally; nothing was competing unnecessarily for attention, and the overall sound was crisp and clean without sounding over-produced.
Test 2: Genre knowledge matters more than genre labels
The second comparison focused on something more difficult than simply following prompts: it tested the models' musical knowledge across multiple genres. Making pop is pretty straightforward, but how well did the models understand different musical traditions? How well could they combine different musical languages in coherent ways?
Case 1: Understanding musical traditions
For this test, I created a prompt based around Indian classical devotional music. Instead of just naming the genre I went into great detail about the musical language: tanpura, bansuri, sitar, harmonium, and pakhawaj, making use of sparse textures and a slow meditative pace to create a peaceful, devotional atmosphere.
While V9 recognized the overall style, it blended many of the elements into a fairly generic "Indian-inspired" sound. The atmosphere was fine, but the arrangement relied more on very recognizable sonic cues rather than on the musical logic behind the tradition.
V9.5 felt much more grounded. The pacing remained patient, the instrumentation fit the devotional mood, and the sparse arrangement delivered the calm atmosphere described in the prompt. It felt that V9.5 didn't just reproduce familiar sounds, but also understood how the different elements worked together.
Listen to the Indian classical devotional music below:
Case 2: Creative fusion that feels intentional
A key test of musical models is how well they can fuse disparate styles together. To test that, I created a prompt for a Country and Future Bass hybrid. The song would begin with a repeating banjo motif before evolving into an exciting and melodic dubstep track.
That kind of prompt requires more than just being able to recognize two genres. It asks the model to develop a single, coherent musical idea that uses two very different stylistic languages.
With V9, the result did incorporate both styles. However, the transition between them felt rather abrupt, like switching between genres rather than developing a single composition that unified them.
V9.5 handled the fusion much more naturally and successfully. The banjo motif wasn't abandoned when the electronic elements entered, and the whole arrangement evolved as one continuous piece rather than two separate ideas awkwardly stitched together.
Taken together, I think these examples suggest that V9.5 isn't just retrieving more musical knowledge. It's better at applying that knowledge in musically coherent ways.
Listen to the Country and Future Bass hybrid below:
Verdict
If you're already using Mureka V9, V9.5 probably won't change the way you write prompts overnight. On straightforward requests, both models are capable of producing polished, usable songs.
The improvements become much more apparent when your prompts require nuance. V9.5 is better at respecting creative intent rather than defaulting to familiar songwriting patterns. Its vocals and mixes feel more natural, and it demonstrates a deeper understanding of musical styles, both in traditional genres and more experimental combinations.
The biggest difference isn't raw audio quality. It's musical judgment.
V9 was already very good at turning prompts into finished songs. V9.5 builds on that, and feels more like a collaborator that understands why those prompts were written in the first place. That's a subtle improvement, but it's also one of the hardest qualities to fake and one that makes music feel noticeably more human.
