Suno.

Suno turns a text prompt into a finished song with vocals, and it has grown well past that. The v6 models (v6 and v6-wild for paying subscribers, v6-mini for everyone) sit alongside a Speech mode for spoken lines, a Sounds mode for effects, remix tools (cover, mashup, extend, replace section, remaster) and Suno Studio, a browser DAW with a chat bar that generates new parts into a multitrack session. Paid subscribers own the songs they make and can use them commercially; songs made on the free plan are for non-commercial use.

Most AI music reviews play one song and give a verdict. We wanted to know what happens when you use Suno like a working project: a small record with a theme, a soundtrack for a real piece of video, and a set of tests where we can measure the result instead of describing it.

So we spent a day on a paid Suno account (upgraded to Premier partway through, for Studio) and built Night Shift, a five-song EP about one night in Singapore. We also used Suno to make every sound in a 30-second hawker-centre film, ran a six-hop "telephone game" of covers, and put the newer tools (Studio's chat bar, Replace Section, Remaster, stems and Custom Models) through specific tasks. The audio is on this page, every prompt is in the appendix word for word, and where a result needs ears rather than a meter, we say so.

What went in, and what came out

Two things in this test came from us rather than from Suno: a silent 30-second film and a 12-second hum. Here they are next to what Suno made from them.

The original film, silent, exactly as we supplied it:

The same film with every sound made in Suno (music, Singlish voices and effects):

The original humming, recorded on a phone:

The result: "Monsoon", the instrumental Suno built from that hum:

The short version

• A tempo in the style box: met on simple prompts. Six of eight precision takes landed within 2.5 BPM by Suno's own tags and our fine-grained tempo analysis, and the other two (a slow waltz) depend on how you count. Busier songs were mixed: Last Train take A came in at 112.6 BPM when we asked for 96 (take B landed at 98), and Hawker Algorithm never switched from 92 to 140 BPM.
• A key in the style box: by Suno's own key tags, two of four prompts matched on both takes, one matched on one take, and one came out in the relative minor (same notes).
• A key change or a beat switch inside a song: not found in any take we measured.
• Lyrics in English: sung choruses were very accurate (0% character error on Last Train's main take). Fast rap was harder: Hawker Algorithm's acronym verse came back with 32% and 75% error.
• Lyrics in Mandarin, Malay and Tamil: Malay was near-perfect, Mandarin was mostly right on the main take, and the Tamil was missing on the main take and only partly there on the other.
• Hits at 0:08, 0:16 and 0:24, then a hard stop at 0:28: the hits arrive, but up to 0.73 seconds early or late. No take stopped at 0:28.
• Extending a song past 6:00: four extensions reached 5:19.
• Replacing one lyric line: the 10 seconds we selected became 35 seconds, and the new line was not sung.
• Studio chat, drums, bass and a lead at 128 BPM: every part sat within 20 ms of the beat grid, but the requested F minor key was ignored.
• Studio chat, an erhu countermelody in the second chorus: it found the chorus by itself, and 92% of its pitched notes were in key.

How we tested

• Everything in Suno, with v6 for every song and Suno's default settings unless we say otherwise (for songs: Weirdness 50%, Style Influence 50%, Variety off, Personalize off). Speech and Sounds used their own defaults.
• Every prompt is published word for word in the appendix at the end.
• Take A is the rule. Suno returns two clips per prompt. We always measured and built on the one Suno lists first. The only exceptions are two spoken lines in the film, which we flag. Each item got at most three attempts, and every attempt counts.
• Measured, not guessed. Lyrics were checked with OpenAI's Whisper large-v3 speech recognizer, often on a vocal track isolated with Demucs. We mostly report character error rate: the share of characters Whisper heard wrong compared with the lyric sheet, so 0% is perfect. Where a table says "words", it is the word error rate instead. Tempo and key came from our own analysis with librosa, and we cross-checked them against the key and BPM tags Suno attaches to each clip.
• Limits of the meters. Whisper is weaker on sung Tamil and Mandarin than on English, key detectors confuse a key with its relative minor or its neighbour a fifth away, and tempo can read as half or double speed. We flag each of these where it matters.

The EP: Night Shift

Suno's Speech mode voiced the intro. The tracks follow it.

1. Last Train: four languages and a key change

The hardest prompt on the record: an 80s city-pop and Mandopop duet at 96 BPM with an English verse, a Mandarin pre-chorus, a Malay verse, a Tamil pre-chorus, a spoken Singlish bridge, and a whole-step key change in the final chorus.

English and Malay came through almost perfectly. In Mandarin, Whisper heard the first line of take A's pre-chorus correctly and two characters of the second line differently (灯火在说话 came back as 等我在说话). We converted Whisper's Traditional-character output to Simplified before scoring. On take A, Whisper found nothing matching the Tamil lines, even when told to expect Tamil. On take B it heard part of the second Tamil line. In take A's final chorus, Whisper heard the Mandarin line 不要睡 ("don't sleep") as the English words "Don't sleep". A note on the table: for short lines, the error score bottoms out when nothing matches, so read scores of about 60% and above as "largely missing".

The tempo request was not met on take A: it runs at 112.6 BPM by Suno's own tag (we measured 112.5). Take B runs at 98. Neither take contains the key change. Take A's key is B♭ major overall. Window by window, our detector also reads F major and D or G minor in places, but never the C major that a whole-step lift would bring. Take B sits in F major.

2. Hawker Algorithm: numbers, acronyms and a beat switch

A rap built to trip up a lyric engine. It packs 21 hard tokens: the year 2026, MRT and HDB, "#08", "1.5 billion", "404", "24/7", "nowrap.ai" and more. The prompt also asks for 92 BPM boom bap that switches to 140 BPM drill halfway through.

• Tokens: take A got 17 of the 21 tokens right. It missed MRT, #08, "char kway teow" and CPF. Whisper wrote "nowrap.ai" as "Now rap A.I.", which we counted as a hit on both takes, because it is how the brand is spoken. Take B got 8 of 21. In take B, Whisper found no rap at all between about 0:45 and 1:25, where the acronym verse should be.
• Beat switch: it never happened. Both takes measure about 139 BPM in both halves (Suno tags take B at 69.6, which is the same speed counted in half time). In effect, the whole song runs at the second half's requested tempo.

3. Monsoon: built from a hum

We hummed a 12-second melody into a phone and uploaded it as the audio input for an instrumental: erhu lead, gamelan bells, Rhodes and rain at 70 BPM. The humming and the finished song are both at the top of this page.

Two things we can measure, and both match. Our key estimate for the humming is G minor, and both Monsoon takes are in G minor. Both run at 70 BPM: we measured 70.0 and 69.8, and Suno tagged them 69.7 and 69.1.

Whether the tune itself survived is a judgement for your ears. We tried two automatic melody comparisons, and neither could tell Monsoon apart from unrelated songs on our list, so we are not claiming a match either way.

4. 3AM Cue: a score with timed hits

A 30-second instrumental for the hawker film below. It asks for hits at 0:08, 0:16 and 0:24, and a hard stop to silence at 0:28, to line up with cuts in the picture. We tried three ways of asking:

1. In words in the style box.
2. The same, plus timed tags in the lyrics box.
3. In bars instead of seconds, at exactly 120 BPM, where bars 5, 9, 13 and 15 fall on the same marks.

For each requested hit, the table shows how far the strongest accent landed from the mark. Negative numbers are early. At 120 BPM, one beat lasts 500 ms.

The hits are there, and they are big: every one sits in the loudest 5% of accents in its take. But they arrive anywhere from 36 to 734 ms off the mark, and no take stops at 0:28. We used attempt 2, take A, for the film, because its three hits are the strongest. We made the stop at 0:28 by hand.

5. First Light: mashup, extend, replace, remaster

The closing track tests the editing tools in a chain. First we mashed up Last Train and Hawker Algorithm with Remix → Mashup, letting Suno merge the lyrics. Then we tried to extend the result past six minutes, replaced one lyric line, and remastered it.

Extend. The mashup came back at 3:33. Our first extension, from the very end of the clip, failed and returned two clips of about 2.5 seconds (Suno refunded the credits). After that we extended from the point Suno suggested each time. Four extensions and "Get Full Song" produced a 5:19 track, so we missed the six-minute target. Suno's metadata shows why the song grew slowly: each extension restarts from an earlier point and drops everything after it. The mashup was cut at 2:28, and the three extensions after it were cut at 0:48, 1:01 and 0:30 of their own length. About two minutes of generated music is not in the final song.

The joins hold up reasonably well. At each of the four join points, we measured how much the sound changes between the two seconds before and the two seconds after. The changes ranked between the 51st and 89th percentile of the song's own section changes, so none was among the most abrupt moments in the song. The key stayed put, and the tempo crept up slightly over the chain, from about 113.7 to 115.5 BPM.

Replace Section. We selected 10 seconds (0:57.5 to 1:07.5) around the first-chorus line "We own this sleepless town" and asked for "Stay awake, Lion City / First light on Marina town". Suno's own metadata shows it received exactly those two lines. Suno returned a 37-second clip, and in the full song the 10 seconds we selected became 35 seconds. The song grew from 5:19 to 5:44.

On an isolated vocal track, Whisper finds "Stay awake, Lion City" repeated several times and then lines borrowed from Hawker Algorithm's hook. "First light on Marina town" is not there, and the original line is gone too. The cut in at 0:57 is one of the most abrupt changes in the whole song by our measure: it ranks in the top 0.3%, with a 6 dB drop in level. Before and after:

Remaster. With the default settings (Normal variation, Boost style), Remaster returned a version of the same length that is very close to the input: 97% waveform similarity over half-second windows, 0.4 dB louder overall, and no frequency band changed by more than 0.6 dB. Key and tempo were unchanged, and on isolated vocals Whisper heard the same sequence of lines in both versions. The player above is the remastered version.

The hawker film: every sound from Suno

The picture is a silent 30-second clip we made separately for this test. It shows a 3 a.m. hawker centre: an uncle calls last orders, a wok flares, a customer orders, the nowrap.ai sign flickers on, rain starts, and the power cuts out. Both versions, silent and scored, are at the top of this article. The music is the 3AM Cue. The seven Singlish lines come from Speech mode, and the wok, sign, thunder, rain and power cut come from Sounds mode.

What we did by hand, so you can judge what Suno did:

• We placed each line and effect on its cue.
• We cut the music at 0:28, where the picture goes black.
• We shortened pauses inside five of the seven spoken lines, removing between 1.0 and 4.3 seconds each.
• For two lines we broke the take A rule. We used take B of "Can, can!", because take A was unintelligible, and take B of the nowrap.ai line, because it was shorter and easier to cut.
• We set the levels, then mixed everything in code (Python and FFmpeg). Suno Studio wasn't used for the film.

Speech mode surprised us. Each script was one short line of dialogue, yet every clip came back 14 to 62 seconds long. Most repeat the line several times, often with laughter in between. Here are the first 15 seconds of "Last order ah! Last order!":

Suno also retitled several clips in other languages: the char kway teow order became 少辣椒的炒粿条, "Can, can! Wait ah, five minute!" became 等五分钟啦, and "Of course lah! AI also need to eat what." became "AI makan tak?". How clearly each line came through:

• Very clear: "Last order ah! Last order!" was heard exactly, and the radio-host intro was 96% right.
• Mostly clear: most lines landed between 12% and 25% error.
• Brand name trouble: "nowrap" kept coming out as "no rap". "Supper meh" came back as "supplement" on take A and "Superman" on take B.
• Unusable: the first take of "Can, can!" was not intelligible to Whisper, so we used take B.

Sounds mode did its job. All six prompts returned two takes each, from 2 to 21 seconds long, and we used five of the six effects in the film.

Telephone Game: six covers in a row

We covered Last Train in a new style, covered that cover, and so on, always building on take A:

1. Bossa nova
2. Metalcore
3. 8-bit chiptune
4. Gospel
5. Lo-fi hip hop
6. Back to the original city-pop style string

Suno's metadata confirms each hop was made from the previous hop's take A. Cover keeps the lyrics box filled with the original lyrics, so lyric survival measures what the model sings, not what it was given. This file plays the first chorus of the original, then the first chorus of each hop in order:

Key and BPM are Suno's own tags. The error columns are for the English chorus and the Malay verse. Melody distance is the average gap in semitones between the first chorus's vocal line and the original's, after lining them up. For scale, choruses with the same words but a different melody (Last Train take B and both Hawker Algorithm hooks) scored 1.73 to 2.14. Three things stand out:

• The key never moved. Every hop stayed in B♭ major.
• The tempo jumped and stayed. The metalcore hop went to about 150 BPM, and every later hop kept that pace, counted in half time as about 74. The last hop's style string still says "96 bpm", yet it came back at 73.4. Cover follows the audio it is given more than the tempo in the prompt.
• The melody went and partly came back. It survived the bossa nova hop clearly, drifted as far as a different melody would be by the metalcore hop, and moved partway back toward the original in the last two hops. The Mandarin line 不要睡, which Whisper never heard on the original, came through clearly on the final city-pop hop.

The lab

Model comparison: v6, v6-wild, v6-mini

We used the same short prompt on all three models: acoustic indie pop at 100 BPM about a morning kopi peng. All six takes ran at 100 BPM. Suno tags most of them as 200, which is the same speed counted in double time.

The players run in table order: v6 A and B, v6-wild A and B, then v6-mini A and B. Suno pitches v6-wild as the less predictable model, and one of its two takes did improvise beyond the lyric sheet. We are running a blind listening test of these six takes. Its result isn't in this article, and we don't rate which model sounds best here.

Precision: tempo and key on demand

We wrote four short instrumental prompts, each asking for a specific tempo and key.

The tempo was right on P2, P3 and P4, within 2.5 BPM on all six takes; those figures are Suno's tags, and our own analysis matched them within 0.2 BPM. P1's slow piano-and-cello waltz depends on how you count it: our analysis puts its takes at 73.6 and 70.6, while halving Suno's double-time tags gives 68.8 and 69.1. Keys are Suno's tags. The key was right on five of eight takes. P2's E♭ minor is the relative minor of F♯ major, the same seven notes with a different home chord, so it is close but not what we asked for. P3's take A came out in A minor, a fifth away. We could not reliably measure the 6/8 waltz feel automatically, so listen to P1 and decide.

Each player plays 15 seconds of take A, then 15 seconds of take B.

Stems: do they add back up to the song?

Studio's automatic split turned Last Train (take A) into 12 stems, from lead vocals and drums to woodwinds. A separation that only splits the mix should add back up to the original almost exactly. These don't.

• The sum misses the master. We added the stems together and subtracted Suno's own WAV of the song. A clean split would leave almost nothing behind. Here, even after matching levels, the leftover is only 8 to 10 dB quieter than the song itself, depending on how the levels are matched. The summed stems also come out about 4 dB louder than the WAV.
• Bass lines up; treble doesn't. Measured over the first 90 seconds, below 150 Hz the stems match the mix closely, with a difference 15 dB down. Above 2 kHz they barely line up at all. That pattern suggests the stems are rendered separately rather than cut out of the finished master.
• The vocal stem is clean. On average it is 27 dB quieter when no one is singing than when someone is.
• A second stems download of the same song is the same split. It came with the same 12 levels, within 0.1 dB, and different file names.
• Advanced split narrows a stem. Using it to pull out just the electric piano gave a stem 3.3 dB lower in level than the automatic "Keyboard" stem, and related to it: their loudness rises and falls together (correlation 0.88).

For remixing, that may not matter. For surgical work, such as muting the vocal while keeping the exact master underneath, don't expect the stems to rebuild the master.

Guardrail: asking for a famous artist's style

We asked for a "pop ballad in the style of Taylor Swift". Suno did not refuse and did not warn us. It quietly saved the style as "pop ballad in the style of Pop, Country, Country Pop, Synth-pop, Indie Folk acoustic guitar heartfelt female vocal" and generated two 17-second clips.

Custom model: does it learn the EP's sound?

Custom Models train a private version of v6 on songs you upload. We uploaded eight EP takes (both takes of the first three songs, the 3AM Cue and the remastered First Light). Suno auto-described each upload in genre terms, and we named the model "Night Shift". Training cost 100 credits and was ready about 11 minutes later. The dialog suggests uploading 24 or more songs for best results, but it let us create the model with eight.

We then ran the same bonus prompt twice, once on the custom model and once on plain v6 as a control: Last Train's style string with a new short lyric.

"Sound distance" compares a rough fingerprint of overall timbre, built from averaged spectral features, with the eight training songs; lower means closer. By that measure the custom-model takes were not closer to the EP than plain v6. They were slightly further away, including from the two Last Train takes that share the style string. It is a crude test, and a model trained on more songs might do better. Listen for yourself:

The players run custom model A and B, then plain v6 A and B.

Suno Studio: an EDM track from a blank session

Studio 2.0 is a browser DAW for Premier subscribers with a chat bar that can generate parts into the session. We started a blank 128 BPM project, imported Last Train for a vocal, and typed the prompts in the appendix one by one. The result, trimmed to its 16 bars:

What we asked the chat bar for, and what it did:

• Drums: a four-on-the-floor beat at 128 BPM for 16 bars. It generated two takes.
• Bass: a rolling bassline in F minor, locked to the kick, with sidechain pumping. It generated two takes, but the bass notes sit on G, E, D, A and B, which is not F minor.
• Lead: a supersaw lead, quiet in bars 1–8 and at full power from bar 9. It generated two takes, but the lead is only about 2 dB louder after bar 9.
• Sidechain compressor on the bass, triggered by the drums: declined, because its compressor has no external sidechain input.
• A custom pumping plugin for the bass: it first described the plugin it could build. After we confirmed, it wrote a plugin called "Quarter Pump" that ducks the volume in time with the beat, saved it to My Plugins and added it to the bass.
• A low-pass filter sweep from bar 5 to bar 9: it added an EQ but said it could not automate the sweep from chat.
• A vocal chop of "Stay awake, Lion City" on bars 9 and 13: the first attempt failed. After a follow-up prompt, it placed both chops and muted the original vocal.

Timing is the strong point. The drums, bass and lead all measure 128.0 to 128.1 BPM. Every strong note we detected starts within 20 ms of a sixteenth-note grid line, with typical offsets of 5 to 10 ms. Our onset detector works in steps of about 6 ms, so that is close to the limit of what we can measure.

Erhu countermelody. In a separate project, we opened Last Train and typed "add an erhu countermelody in the second chorus". The chat bar located the second chorus at bars 44–53 by itself and asked us to confirm. It then wrote its own style prompt for the part and generated two takes. Take 1, which we kept, starts 0.14 seconds after the second chorus vocal comes in and covers the whole chorus. 92% of its pitched notes fall inside B♭ major, the song's key. It sits about 5 dB under the mix.

Smaller things we noticed in Studio:

• Open in Studio quietly time-stretched the song. It shortened Last Train from 3:23.2 to 3:22.3, or 0.4%, to fit the project's 113 BPM grid.
• Exports are free to download. Full-song exports land in your library, and downloading them didn't use any of our download allowance.
• A WAV export stalled. It sat on "Preparing…" for more than four minutes.
• Mutes are handled inconsistently. The multitrack ZIP included a track we had muted, at full level. Our 30-second Hawker Rave also exported as a 3:14 file, because that muted track ran on past bar 16.

What it cost

• Credits: 580 of the account's 10,005 credits. A song prompt costs 10 credits for two takes, and the custom model cost 100.
• Downloads: 55 of the 61 song downloads available on the account. Premier caps song downloads per month, while Studio exports are exempt.

Who it is for

For marketers and freelancers who need a finished-sounding bed for a short video, Suno turns a prompt into a full, mixed song in a couple of minutes, and our hawker film went from silent to fully scored using nothing but Suno audio. The same goes for idents, pitches and scratch tracks you will later replace. If your work is on the marketers' desk, that is the realistic use. Plan to do the structural work yourself: timing, song structure and precise edits are where it fell short in our tests.

If you are comparing AI tools for creative work, our Blender test of two coding models took the same approach: one prompt, real output, and the files next to the verdict.

Appendix: every prompt, word for word

Defaults unless noted: model v6, Weirdness 50%, Style Influence 50%, Variety off, Personalize off, Duration Auto.

Last Train: style:

80s city pop meets Mandopop, 96 bpm, warm Rhodes, slap bass, gated drums, analog synth pads, male and female duet, glossy night-drive mood, final chorus modulates up a whole step

Lyrics:

[Intro]

[Verse 1 - Male, English]
Last train sliding out of Raffles Place
Neon on the window, sleep upon my face
Half the island's dreaming, half is clocking in
I'm counting every station 'til the night begins

[Pre-Chorus - Female, Mandarin]
夜色那么长，我们不回家
灯火在说话

[Chorus - Duet]
Stay awake, Lion City
Don't let the lights go down
Stay awake, Lion City
We own this sleepless town

[Verse 2 - Female, Malay]
Angin malam di tepi Marina
Kita berjalan tanpa arah
Bandar ini tak pernah lena
Hati kita masih menyala

[Pre-Chorus - Male, Tamil]
இரவு இன்னும் முடியவில்லை
நம் கனவுகள் விழித்திருக்கின்றன

[Chorus - Duet]
Stay awake, Lion City
Don't let the lights go down
Stay awake, Lion City
We own this sleepless town

[Bridge - Spoken, Singlish]
Eh bro, last train already gone lah.
Never mind lah, we walk. Can one.

[Key Change]

[Final Chorus - Duet]
Stay awake, Lion City
不要睡
Stay awake, Lion City
We own this sleepless town

[Outro - Solo Rhodes, fade out]

Hawker Algorithm: style:

Singapore hip hop, rapid-fire male rap, 92 bpm boom bap in the first half, beat switch to 140 bpm drill in the second half, kopitiam ambience, crisp diction

Lyrics:

[Intro - Spoken]
Nowrap Dispatch, 3 a.m. edition.

[Verse 1 - Rap]
2026, still queueing at the stall
GPT on my phone but the uncle knows it all
HDB to MRT, CBD by 4
17 tabs open and I'm still scrolling more
Plate #08, char kway teow with the hum
nowrap.ai on the sign, tell me what's to come

[Hook]
Stay awake, Lion City
Algorithm on a hawker plate
Stay awake, Lion City
Data hot and the night runs late

[Beat Switch - 140 bpm drill]

[Verse 2 - Rap, double time]
API, LLM, RAG on a GPU
1.5 billion tokens and a kopi for two
CPF, COE, every acronym in the queue
404 on my sleep but the server's pushing through
5G at the void deck, 24/7 neon glow
3 a.m. economy and the uncle's on the go

[Hook]
Stay awake, Lion City
Algorithm on a hawker plate
Stay awake, Lion City
Data hot and the night runs late

[Outro - Spoken]
Uncle, one more plate!

Monsoon: humming uploaded with "+ Audio" (Cover, Audio Influence 25%); style:

instrumental, monsoon night, erhu lead carrying the hummed melody, gamelan bells, warm Rhodes, soft rain textures, slow 70 bpm, cinematic

3AM Cue: instrumental, Duration Custom 0:30. Style for attempts 1 and 2:

instrumental cinematic hybrid score, 120 bpm, 4/4, tense late-night hawker centre; quiet pulsing intro; huge taiko and brass hit at 0:08; bright synth swell and hit at 0:16 as a neon sign switches on; thunderous hit at 0:24 into driving percussion; abrupt hard stop to total silence at 0:28

Attempt 2 lyrics box:

[Intro 0:00 - soft pulse]
[Hit 0:08 - taiko and brass]
[Hit 0:16 - synth swell]
[Hit 0:24 - thunder hit, driving drums]
[Hard stop 0:28 - silence]

Attempt 3 style and lyrics box:

instrumental cinematic hybrid score, exactly 120 bpm, 4/4, tense late-night hawker centre; bars 1-4 quiet pulsing intro; huge taiko and brass hit on the downbeat of bar 5; bright synth swell and hit on the downbeat of bar 9; thunderous hit on the downbeat of bar 13 into driving percussion; abrupt hard stop on the downbeat of bar 15, then total silence

[Bars 1-4: quiet pulse]
[Bar 5 downbeat: HUGE taiko and brass hit]
[Bar 9 downbeat: synth swell hit]
[Bar 13 downbeat: thunder hit, driving drums]
[Bar 15 downbeat: hard stop]
[Silence]

First Light: Remix → Mashup of Last Train (take A) and Hawker Algorithm (take A), with Suno's "Mashup Lyrics" merge; style:

80s city pop meets Singapore hip hop, warm Rhodes, slap bass, boom bap drums, male and female duet with a male rapper, glossy night-drive mood, building to a sunrise finale

Extension lyrics, in order (same style):

[Bridge - Duet]
First light over Marina Bay
Kopi steam as the sky turns gold from grey

[Rap - singer B]
6 a.m., the MRT doors slide
Uncle pulls the shutter, the night says goodbye

[Final Chorus - Duet]
Stay awake, Lion City
The sun is coming round
Stay awake, Lion City
We own this sleepless town

[Outro - fade out]

Stay awake, Lion City
We own this sleepless town

[Instrumental break - Rhodes and slap bass solo]

[Verse 3 - singer A]
Taxi lights are fading on the PIE
Hawker smoke is drifting to the morning sky

[Rap - singer B]
Kopi-O kosong, 6:15
Whole city rebooting like a new machine

[Final Chorus - Duet]
Stay awake, Lion City
First light coming down
Stay awake, Lion City
We own this sleepless town

[Outro - solo Rhodes, fade out]

Stay awake, Lion City
We own this sleepless town

[Breakdown - Spoken, Singlish]
Eh, sun coming up already leh.
Kopi one more round, then we go home.

[Verse 4 - singer A, soft]
Shutters rising on the shophouse row
Last of the neon giving up its glow

[Rap - singer B]
Night shift over but the beat don't quit
Save the file, close the laptop, that's it

[Final Chorus - Duet, big]
Stay awake, Lion City
First light coming down
Stay awake, Lion City
We own this sleepless town

[Long instrumental outro - Rhodes, slap bass, city ambience, fade out]

[Reprise - Duet]
Stay awake, Lion City
Stay awake, Lion City
First light coming down
We own this sleepless town

[Instrumental - Rhodes solo over slap bass groove]

[Final Hook - Duet]
Stay awake, Lion City
We own this sleepless town

[Outro - city ambience, slow fade out]

The first extension's lyrics were used twice: on the failed attempt from the very end of the clip, and on the retry from Suno's suggested point, where Suno prefilled the box with the lyrics leading into it. Replace Section, 0:57.5 to 1:07.5:

Stay awake, Lion City
First light on Marina town

Remaster: default settings (Normal variation, Boost style).

Hawker film: Speech mode (background music off). Each line is voice, then tone field, then script:

V1 Male | older Singaporean hawker uncle shouting across an empty hall, Singlish accent | Last order ah! Last order!
V2 Male | young Singaporean customer, casual, Singlish accent | Uncle, char kway teow one plate, tapau! Less chilli ah!
V3 Male | older Singaporean hawker uncle, busy and cheerful, Singlish accent | Can, can! Wait ah, five minute!
V4 Female | young Singaporean woman, amused, Singlish accent | Eh, nowrap.ai also sell supper meh?
V5 Male | young Singaporean man, laughing, Singlish accent | Of course lah! AI also need to eat what.
V6 Female | young Singaporean woman, startled by sudden rain, Singlish accent | Alamak, rain liao! Quick, quick!
V7 Male | older Singaporean hawker uncle, startled in the dark, Singlish accent | Eh, who off the light?!
Intro Male | calm late-night radio host, Singaporean English accent | This is the Nowrap Dispatch. Night Shift: five songs, one night in Singapore, made with Suno.

Sounds mode:

single loud thunder crack close by, then a rolling rumble
heavy tropical rain on a zinc roof, steady   (loop)
wok flame burst whoosh with sizzling oil
fluorescent lightbox sign flickering on, two electrical buzz-clicks then a steady hum
power cut, electrical thunk and ceiling fans spinning down to silence
train departing a platform at night

Telephone Game (Remix → Cover, Audio Influence 25%, always from the previous hop's take A). Styles:

bossa nova, nylon guitar, brushed drums, soft male and female vocals
metalcore, drop-tuned guitars, breakdowns, screamed and clean vocals
8-bit chiptune, NES square waves, retro game soundtrack, vocoder vocals
gospel choir, Hammond organ, handclaps, call and response
lo-fi hip hop, dusty vinyl, mellow keys, soft sung vocals
80s city pop meets Mandopop, 96 bpm, warm Rhodes, slap bass, gated drums, analog synth pads, male and female duet, glossy night-drive mood, final chorus modulates up a whole step

Model comparison (same prompt on v6, v6-wild and v6-mini):

acoustic indie pop, 100 bpm, ukulele, light percussion, warm female vocal, sunny morning kopitiam

[Verse]
Kopi peng in the morning heat
Uncle knows my order, two sugar, no repeat

[Chorus]
Ice is melting, day is starting
Kopi peng, my heart is smiling

Precision:

instrumental, solo piano and cello, 72 bpm, A minor, 6/8 waltz
instrumental, deep house, four-on-the-floor, 124 bpm, F# major
instrumental, liquid drum and bass, 174 bpm, D minor
instrumental, funk, slap bass and horns, 96 bpm, B-flat major

Guardrail:

pop ballad in the style of Taylor Swift, acoustic guitar, heartfelt female vocal

[Verse]
Window seat on the midnight bus
city lights keep calling us

Custom model bonus (custom model "Night Shift", then plain v6 with the same settings); style as Last Train, lyrics:

[Verse]
Paper cup of kopi, the night bus running late
Neon on the river, every shutter painted grey

[Chorus]
Night shift, Lion City
We keep the lights on for you
Night shift, Lion City
Till the morning breaks through

Studio chat bar, Hawker Rave (blank project, 128 BPM, in order):

Add a punchy four-on-the-floor EDM drum beat at 128 bpm for 16 bars: big kick, claps on 2 and 4, open hi-hats on the offbeats
Add a rolling EDM bassline in F minor for the same 16 bars, locked to the kick, with sidechain pumping
Add a bright supersaw lead synth playing a catchy hook in F minor, quiet in bars 1-8 and full power from bar 9 as the drop
Put a sidechain compressor on the Bass track, triggered by the Drums track, so the bass ducks on every kick
Design a custom pumping plugin for the Bass track that ducks the volume on every quarter note at 128 bpm, like classic EDM sidechain
Build the tempo-synced pumping effect
Put a low-pass filter on this track and automate it: closed at bar 1, sweeping slowly open through bars 5 to 8, fully open from bar 9 for the drop
Turn this vocal into an EDM vocal chop: keep only the sung phrase 'Stay awake, Lion City' and place it on the downbeats of bar 9 and bar 13, removing the rest of the vocal
Yes, go ahead: isolate the phrase from the existing vocal and place copies on the downbeats of bars 9 and 13

Studio chat bar, erhu (Last Train take A opened as a single track):

add an erhu countermelody in the second chorus

Sources: Suno — Introducing v6, Suno — Introducing Studio 2.0, Suno — v5.5 and Custom Models, Music Business Worldwide — Suno download limits, Suno Help — Does Suno own the music I make?