What happened
Artificial Analysis’s AA-Music-Vocal v1.1 leaderboard places ACE Music v1 at rank one, followed by Suno v6 and Suno v6-mini. Their listed Elo scores are 1160, 1144 and 1132, respectively. The board also includes Mureka V9.5 at 1101 and Mureka V9 at 1063. Google’s Lyria 3.5 and Lyria 3 Pro are listed at 1050 and 1048; Eleven Music v2.5 is at 1016. The leaderboard is specifically for models generating music with singing or other vocal elements. Artificial Analysis keeps a separate board for instrumental music and notes that systems may perform differently across the two.
The ranking method is based on blind pairwise comparisons. Artificial Analysis says evaluators listen to two music samples generated from the same prompt and choose a preferred track without knowing which model produced it. The company describes the evaluators as validated private evaluators. Scores use an Elo system, with higher values indicating that a model is preferred more often. MiniMax Music 3.0 is anchored at 1000 as the reference point on this board’s scale.
The displayed sample counts differ: ACE Music v1 has 2,119 samples, while Suno v6 has 3,063 and Suno v6-mini has 3,175. The leaderboard gives ACE Music v1 a rank range of 1–2 and an Elo interval from 1139 to 1181. Suno v6 has a rank range of 1–3 and an interval from 1126 to 1162. Those intervals overlap, so the listed scores do not establish a clear statistical separation between the top two. Artificial Analysis’s FAQ confirms the top five in the same order and says the rankings are based on blind votes.
The rest of the published table shows how far the scale extends below the leading group. Eleven Music v2 is listed at 1002, MiniMax Music 3.0 at 1000, Eleven Music at 998, MiniMax Music 2.6 at 998, and MiniMax Music 2.5+ at 990. Sonauto Melodia v3 occupies the final listed position at 964. Several entries have broad rank ranges, including Eleven Music v2 at 8–13 and MiniMax Music 3.0 at 8–13, which reinforces that a displayed ordinal rank is not the same as a definitive separation from every nearby system.
The page does not describe ACE Music v1’s release date, availability, pricing, licensing, or technical architecture. It reports a benchmark position rather than announcing a product change. Builders evaluating music generation can treat this as a candidate shortlist signal, while recognizing that the rank reflects the preferences captured by this particular evaluation setup.
What the archive adds
New Runtime’s earlier coverage reported that ACE-Step published ACE-Step 1.5 XL, a 4B DiT variant of its local music-generation project (ACE-Step 1.5, 2026-10-24). That is related by ACE’s music-generation work, but the current leaderboard result is a new fact, not evidence that ACE Music v1 is the same system as ACE-Step 1.5 XL.
Trend signal
This result does not yet fit a pattern New Runtime tracks. It is a bounded music-model leaderboard update rather than evidence about application infrastructure, agent interoperability, verification workflows, or the production boundary for AI systems. Its immediate value is comparative evaluation, not a demonstrated shift in how builders deploy or govern generative media.
Why it matters
For teams building music products, the leaderboard can help identify which systems deserve hands-on evaluation when generated vocals matter. ACE Music v1’s first-place score is a useful signal, but the small difference from Suno v6 and the overlapping intervals argue against treating the order as a decisive purchasing or integration rule.
A team should test candidates with prompts representative of its own genres, languages, vocal styles, and production constraints. Blind comparison is also a useful method to borrow internally: remove model labels, keep prompts constant, and have listeners judge outputs against the product’s real criteria. This turns a public ranking into a selection process tied to the experience a product actually needs to deliver.
The benchmark’s rankings cannot answer questions about price, latency, commercial rights, editability, API access, or reliability. Those factors may matter more than a preference score once a team must generate music repeatedly inside a product. A leading rank is therefore most useful as an input to evaluation design: put ACE Music v1 and the closest alternatives into the same controlled test rather than replacing product-specific assessment with the public table.
Limits and uncertainty
Artificial Analysis describes the pairwise method and lists sample counts, but the leaderboard page does not explain how its validated private evaluators are recruited, how prompts are selected, how votes are weighted, or how often the table is updated. It does not provide enough information here to independently reproduce the ranking or determine whether the samples reflect particular genres, languages, vocal styles, or production conditions.
The intervals for ACE Music v1 and Suno v6 overlap, and the scores measure evaluator preference rather than objective fidelity, controllability, lyric quality, safety, or suitability for a particular product. The comparison also does not establish why listeners preferred one output. Availability, pricing, licensing, technical access, and model architecture are not established by this ranking.
Reader takeaway
ACE Music v1 leads Artificial Analysis’s current vocal-music preference board, but the narrow margin and overlapping intervals make this a reason to test it, not to assume it is universally best. Use blind, prompt-matched listening tests for your own use case before committing to a generator.