AI music is moving fast, and ElevenLabs Music v2 is one of the most ambitious new song-generation models. But can it preserve supplied lyrics, follow a detailed arrangement, and deliver a usable track without heavy editing?
We tested it with three original 90-second briefs: a synth-pop vocal song, a cinematic instrumental, and an eight-part alternative-pop arrangement. This review covers the audio, prompt compliance, generation time, task-reported cost, official controls, and API access.
Quick Answer: How Good Is ElevenLabs Music v2?
OUR SHORT VERDICT
ElevenLabs Music v2 was strongest when the brief demanded explicit lyrics and section control. It produced all three requested 90-second files, kept the supplied words intact in our lyric test, and made the eight requested song sections easy to hear.
It was less exact at the edges: one instrumental impact arrived earlier than requested, one instrumental break ran long, and the structured song ended more like a fade than the clean final chord in the prompt.
Best observed strength: lyrics and audible song structure.
Main limitation: fine timing and ending instructions were not always exact.
Test scope: three controlled API runs, not an industry benchmark.
ElevenLabs Music v2 is also available through GlobalGPT’s audio generator. That route puts multiple audio and AI models in one account, avoids a separate VPN or provider setup, and can reduce the cost of maintaining several standalone subscriptions.
Before looking at our results, it helps to separate the model’s published feature set from what three prompts can actually prove.
什麼是 ElevenLabs Music v2？
ElevenLabs introduced Music v2 on May 26, 2026 as an upgraded music-generation model for songs and instrumentals. The company says the release improves vocals, instrumentation, arrangement, multilingual support, dense lyric delivery, genre transitions, and inpainting.
Music v2 powers ElevenMusic, ElevenAPI, and ElevenCreative. Through the API, developers can generate tracks programmatically, edit a selected region with inpainting, and work with reference matching.
MODEL ID music_v2
MAXIMUM DURATION 5 minutes
PUBLISHED API RATE $0.15/minute
AUDIO OUTPUT 44.1 kHz
PUBLISHED BITRATE 128–192 kbps
API ACCESS 付費用戶
ElevenLabs’ official Music v2 release page, checked August 2026.
ElevenLabs Music v2 Test Results
We used three controlled API runs with a fixed 90-second target. GPT generated the original lyric sets once on our platform, and we froze them before testing Music v2. Listening notes were cross-checked with Gemini 3.6 Flash audio analysis. Times below are local end-to-end generation time; charges are values reported by the completed tasks.
Test 1: Pop Song With Supplied Lyrics
TEST 01Pop Song With Supplied Lyrics
90 秒目標每分鐘 112 次未成年人女高音
任務設定： Create a contemporary synth-pop song about two people leaving a city late at night, using the supplied words exactly. We listened for lyric preservation, vocal clarity, hook strength, arrangement, and the requested final chord.
Test 01 resultStrong lyric consistency with a longer electronic intro
90.0 秒輸出
42.0 秒GENERATION TIME
$0.225TASK CHARGE
歌詞並無明顯偏差 was detected in the supplied words.
The output had a polished electronic synth-pop character and clean vocal/instrument separation.
The intro felt longer than requested before the vocal arrived.
Test 2: Cinematic Instrumental
TEST 02Cinematic Instrumental
90 秒目標每分鐘 82 次D小調無人聲
任務設定： Build a science-fiction cue around a recognizable felt-piano motif, with gradual orchestral growth, brass only in the final third, no human voice texture, and a resolved ending.
Show the full copyable prompt
為一幕充滿戲劇張力的科幻探索場景，創作一首完全原創、長度為 90 秒的電影風格器樂配樂。節拍為 82 BPM，調性為 D 小調。不得包含人聲、對白、低語、合唱或任何類似人聲的音效。 以稀疏的軟質鋼琴主題和低音弦樂的長音靜謐開場。開場後引入微妙的類比節拍，隨後透過層疊的弦樂與克制的低音打擊樂逐步鋪陳。僅在最後三分之一處加入銅管樂器。接近結尾時達到一個清晰的情感高潮，但避免過度響亮或失真。 保持鋼琴主題的辨識度，並確保鋼琴、弦樂、打擊樂、合成器脈動與銅管樂器之間有清晰的分離度。採用真實的管弦樂音色，混音效果應寬廣但受控。前半段應避免使用預告片常見的衝擊音效。請勿模仿現有的配樂、作曲家或系列作品。結尾應以一個解決後的長音和弦收尾，而非淡出。.
Test 02 resultA convincing cinematic build, with one early impact
90.0 秒輸出
20.9 秒GENERATION TIME
$0.225TASK CHARGE
No voice-like texture was detected.
The piano motif remained recognizable as the arrangement built around it.
An impact arrived earlier than the brief allowed, weakening strict timing compliance.
Test 3: Eight-Part Song Structure
TEST 03Eight-Part Song Structure
90 秒目標每分鐘 104 次D大調固定章節順序
任務設定： Create an alternative-pop song with eight audible sections in a fixed order: Intro, Verse, Pre-Chorus, Chorus, Instrumental Break, Bridge, Final Chorus, and Outro.
Test 03 resultThe clearest structural result, but not perfect timing control
90.0 秒輸出
20.9 秒GENERATION TIME
$0.225TASK CHARGE
All eight sections were easy to identify, with strong bridge and final-chorus contrast.
The instrumental break ran longer than the requested eight seconds.
The ending behaved more like a fade than the clean final chord in the prompt.
What the Three Tests Show
Performance Findings
3/3Usable outputs
Every run returned a playable 90-second file that fit the broad musical task.
1Clear lyric pass
The supplied pop lyrics showed no clear deviation in our listening checks.
8Audible sections
The fixed-structure song made all requested sections easy to identify.
Vocals: clear, centered, and separated well from the electronic backing in Test 1.
Instrumental writing: the piano motif and controlled build in Test 2 showed that the model can handle more than vocal pop.
Prompt control: broad structure was stronger than frame-level timing; intros, breaks, impacts, and endings still deserve manual review.
Editing burden: none of the three outputs was unusable, but Tests 2 and 3 would need small arrangement edits for strict production delivery.
Efficiency Findings
Generation Time Across Three 90-Second Runs
Pop vocals
42.0 秒
Instrumental
20.9 秒
Eight-part song
20.9 秒
Average local end-to-end generation time: 27.9 seconds. Local time includes request handling, polling, and network delay; it is not provider-side latency.
Total output: 270 seconds across three files.
Total task-reported charge: $0.675.
Charge per run: $0.225 for 90 seconds, which equals 1.5 minutes × the official $0.15-per-minute API rate.
No reruns: all three completed tasks produced usable audio on the first recorded attempt.
How to Access ElevenLabs Music v2
Access Through ElevenLabs
Use ElevenMusic for browser-based composition.
Use ElevenAPI for programmatic generation and editing.
The official SDK exposes the model through elevenlabs.music.compose 與 model_id="music_v2". A basic request can set a natural-language 提示 和 music_length_ms. Composition plans add more precise section structure, lyric timing, and arrangement control.
Generation: create a complete song or instrumental from a prompt.
Detailed response: request more information about the generated composition.
串流： begin receiving audio without waiting for the complete file.
Inpainting: regenerate a selected region instead of replacing the whole track.
Reference matching: guide a generation with reference material where supported.
Prompts that name copyrighted artists, songs, or protected lyrics can trigger a bad_prompt 或 bad_composition_plan response. Describe musical attributes instead of asking for a living artist’s exact style.
Example API Request
Python SDKModel: music_v2
import os
from elevenlabs import ElevenLabs, save
client = ElevenLabs(api_key=os.environ["ELEVENLABS_API_KEY"])
audio = client.music.compose(
model_id="music_v2",
prompt=(
"Create an original cinematic instrumental with felt piano, "
"low strings, a gradual build, and a resolved final chord."
),
music_length_ms=90_000,
)
save(audio, "music-v2-test.mp3")
Keep the API key in an environment variable, not inside the script. For a production app, add request logging, retry limits, and a clear cost cap before generating long tracks.
Maximum published output length for one music request.
Our 90-second task charge
$0.225
1.5 minutes × $0.15; reported by each completed test task.
Our three-test total
$0.675
Three 90-second outputs; no rerun cost included.
The unit rate in our test record matches the official API rate. GlobalGPT’s value here is consolidated access and fewer separate subscriptions—not a verified lower per-minute Music v2 price.
How to Prompt ElevenLabs Music v2
Music v2 responded best when the brief separated the musical goal from the non-negotiable constraints. Build your prompt in this order:
1輸出Song or instrumental, plus duration.
2Musical frameGenre, mood, tempo, key, and vocal type.
3ArrangementOpening, section changes, climax, and ending.
4Mix prioritiesVocal position, separation, loudness, and space.
5除外事項No speech, no fade, no crowd, or no voice texture.
6Lyrics and structureSection labels and whether words may change.
Useful English Music Terms
Click any term or descriptor to copy it. 範例： warm, intimate lead vocal with clear diction.
