How long can an AI song be? Suno, ElevenLabs and Lyria compared

ElevenLabs documents Music lengths of 3 seconds to 5 minutes. Suno v6 is reported at up to eight minutes by a third-party site (I could not read Suno's own announcement, so treat that as unconfirmed), and Google says Lyria 3.5 makes full songs of about a couple of minutes, controllable through the prompt. If you need longer, generate sections and join them.

Side by side

Sume has no duration field

The Sume Music Router rejects duration and duration_seconds. You shape length by writing section markers such as [0:00-0:30] Intro: in the prompt, as the router docs describe.

curl -X POST https://api.sume.com/v1/music-router/generate \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: long-song-001" \
  -d '{
    "model": "sume/music-auto",
    "prompt": "Cinematic orchestral build, 90 BPM, D minor. [0:00-0:30] Intro: low strings. [0:30-1:30] Build: brass and timpani. [1:30-2:00] Resolve: solo cello. Instrumental."
  }'

Joining tracks

Timeline audio concatenates 1 to 20 parts into one file up to 1,800 seconds for a flat $0.01 per job. It joins at the sample level without re-synthesis. Parts must be audio on media.sume.com, and an Idempotency-Key is required.

Example, with an assumption stated: if four generations each return about two minutes, the cost is 4 x $0.125 + $0.01 = $0.51 for roughly eight minutes.

Caveat

• Concat joins audio; it does not match key or tempo between parts. Brief all parts with the same BPM and key.
• The router price is per generation, not per minute.
• Check the length of every part before joining.