Seedance 2.5 has third-party benchmark data, but only on one of the two major public arenas. Arena.ai's Video Arena scores it in all three video categories: 6th in Text-to-Video, 4th in Image-to-Video, and 2nd in Video Edit. Artificial Analysis's Video Arena has no Seedance 2.5 row at all, and its newest ByteDance entry is still Seedance 2.0 720p. VBench has never listed a Seedance model of any version. That split matters, because most "Seedance 2.5 benchmark" claims online quietly mix numbers from boards that use different voters, different prompts, and score scales that are not interchangeable. What follows is every public score verifiable today, with the source board, that board's own update date, the vote count behind each number, and what each leaderboard actually measures. The Veo, Kling, and Runway comparisons the raw tables support are here too, along with the ones they do not support.
Last verified: 2026-09-17. Every score below was read directly from the live leaderboards on that date.
TL;DR
- Arena.ai Video Arena is the only major public board scoring Seedance 2.5 today: 6th in Text-to-Video (1482), 4th in Image-to-Video (1475), 2nd in Video Edit (1410).
- Seedance 2.5 outscores every Veo, Kling, and Runway entry on all three Arena.ai video boards. The closest Veo is
veo-3.1-audioat 12th (1364) in Text-to-Video; the closest Runway isrunway-gen-4.5at 27th (1224). - Artificial Analysis has no Seedance 2.5 row on its Text-to-Video, Image-to-Video, or Video Editing leaderboards. An "Artificial Analysis Seedance 2.5 Elo" is a number that is not on the board.
- Seedance 2.5 is a statistical tie with Seedance 2.0 on both generation boards. The gaps are 3 and 1 points against confidence intervals of ±12 and ±8. Video Edit is where 2.5 pulls clearly ahead.
- The features 2.5 was built around, including 30-second single-pass clips, 50 reference inputs, and timestamp-level edits, are not what any public arena votes on. Arena votes land on short generations at default settings.
Quick Answer
On Arena.ai's Video Arena, dreamina-seedance-2.5-720p scores 1482 ±12 in Text-to-Video (rank 6 of 48), 1475 ±8 in Image-to-Video (rank 4 of 48), and 1410 ±26 in Video Edit (rank 2 of 10). Those are blind human-preference scores rather than quality percentages, and they only mean something within the board that produced them. Seedance 2.5 has no Artificial Analysis score and no VBench score, because neither has added the model. For what the model does rather than how it gets voted on, the Seedance 2.5 model page lists its current capabilities and limits.
Arena.ai Text-to-Video Arena, board dated Sep 4, 2026, captured 2026-09-17. Source: Arena.ai Text-to-Video Arena
Every public Seedance 2.5 score in one table
| Leaderboard | Category | Seedance 2.5 result | Board's own date | Votes behind it |
|---|---|---|---|---|
| Arena.ai Video Arena | Text-to-Video | 6th of 48, score 1482 ±12 (rank spread 3 to 9) | Sep 4, 2026 | 4,066 |
| Arena.ai Video Arena | Image-to-Video | 4th of 48, score 1475 ±8 (rank spread 2 to 6) | Sep 14, 2026 | 12,151 |
| Arena.ai Video Arena | Video Edit | 2nd of 10, score 1410 ±26 (rank spread 1 to 3) | Aug 26, 2026 | 429 |
| Artificial Analysis Video Arena | Text-to-Video (With Audio) | Not listed | n/a | n/a |
| Artificial Analysis Video Arena | Image-to-Video (With Audio) | Not listed | n/a | n/a |
| Artificial Analysis Video Arena | Video Editing | Not listed | n/a | n/a |
| VBench | Text-to-Video quality dimensions | Not listed, and no Seedance version ever has been | n/a | n/a |
Two columns there do more work than the headline ranks. Rank spread is the range of positions a score could occupy once its confidence interval is applied, so Seedance 2.5's Text-to-Video spread of 3 to 9 is the board itself declining to separate the model from the five above it and the three below. Vote count is the other one. Seedance 2.5 carries 4,066 Text-to-Video votes against Seedance 2.0's 52,864, which makes its score the most likely number on this page to move before the next update.
Text-to-Video: Seedance 2.5 against Veo, Kling, and Runway
| Rank | Model | Lab | Arena Score | Votes |
|---|---|---|---|---|
| 1 | gemini-omni-1.1-flash |
1515 ±15 | 1,777 | |
| 3 | wan3.0 |
Alibaba | 1494 ±19 | 1,167 |
| 6 | dreamina-seedance-2.5-720p |
ByteDance | 1482 ±12 | 4,066 |
| 7 | dreamina-seedance-2.0-720p |
ByteDance | 1479 ±8 | 52,864 |
| 11 | sora-2-pro |
OpenAI | 1367 ±7 | 50,415 |
| 12 | veo-3.1-audio |
1364 ±14 | 13,703 | |
| 22 | seedance-v1.5-pro |
ByteDance | 1256 ±7 | 75,891 |
| 27 | runway-gen-4.5 |
Runway | 1224 ±9 | 43,610 |
| 28 | kling-2.5-turbo-1080p |
KlingAI | 1219 ±17 | 2,100 |
| 29 | kling-2.6-pro |
KlingAI | 1216 ±7 | 73,062 |
Arena.ai Text-to-Video Arena, board dated Sep 4, 2026, read 2026-09-17. Selected rows from a 48-model board.
The supportable reading is narrow: on blind text-prompt votes, Seedance 2.5 sits in the leading group, about 118 points above the best-scoring Veo entry and about 258 above the best-scoring Runway entry. What the table does not support is a sentence like "Seedance 2.5 beats Veo 3.1 by 118 points on video quality." Voters compared whole generations from single prompts with audio in play, so one number absorbs motion, sound, prompt adherence, and personal taste at once. To see the kind of job these votes were cast on, the AI text-to-video generator runs the same class of short single-prompt request.
Image-to-Video: the board with the most votes behind it
Arena.ai Image-to-Video Arena, board dated Sep 14, 2026, captured 2026-09-17. Source: Arena.ai Image-to-Video Arena
| Rank | Model | Lab | Arena Score | Votes |
|---|---|---|---|---|
| 1 | minimax-h3 |
MiniMax | 1494 ±5 | 49,443 |
| 2 | gemini-omni-1.1-flash |
1488 ±11 | 3,731 | |
| 3 | wan3.0 |
Alibaba | 1479 ±11 | 3,444 |
| 4 | dreamina-seedance-2.5-720p |
ByteDance | 1475 ±8 | 12,151 |
| 5 | dreamina-seedance-2.0-720p |
ByteDance | 1474 ±7 | 118,475 |
| 13 | veo-3.1-audio |
1398 ±10 | 25,116 | |
| 19 | kling-v3-pro |
KlingAI | 1355 ±6 | 199,238 |
| 26 | kling-2.6-pro |
KlingAI | 1294 ±8 | 202,157 |
| 47 | runway-gen4-turbo |
Runway | 1052 ±13 | 6,801 |
Arena.ai Image-to-Video Arena, board dated Sep 14, 2026, read 2026-09-17. Selected rows from a 48-model board.
This board carries the largest sample of the three, 2,013,261 votes across all models, and it answers the Seedance 2.5 versus 2.0 question more cleanly than anything else available. One point apart, against intervals of ±8 and ±7. On generic single-image animation prompts the two versions are indistinguishable, which says more about what the board measures than about either model. Your own still, run through the AI image-to-video generator, will tell you more about your footage than a one-point gap ever will.
Video Edit: where 2.5 separates from 2.0
Arena.ai Video Edit Arena, board dated Aug 26, 2026, captured 2026-09-17. Source: Arena.ai Video Edit Arena
| Rank | Model | Lab | Arena Score | Votes |
|---|---|---|---|---|
| 1 | wan3.0 |
Alibaba | 1414 ±26 | 463 |
| 2 | dreamina-seedance-2.5-720p |
ByteDance | 1410 ±26 | 429 |
| 3 | minimax-h3 |
MiniMax | 1392 ±19 | 962 |
| 5 | dreamina-seedance-2.0-720p |
ByteDance | 1365 ±14 | 3,852 |
| 8 | kling-o3-pro |
KlingAI | 1255 ±11 | 7,540 |
| 9 | kling-o1-pro |
KlingAI | 1197 ±10 | 10,542 |
| 10 | runway-gen4-aleph |
Runway | 1182 ±9 | 9,818 |
Arena.ai Video Edit Arena, board dated Aug 26, 2026, read 2026-09-17. The full 10-model board. No Veo model is entered in this category.
Video Edit is the newest and smallest category, 10 models and 27,930 votes in total, and it is the one place where Seedance 2.5 opens a gap on its predecessor that survives the confidence intervals: 45 points clear of Seedance 2.0, and 228 clear of Runway's Gen-4 Aleph at the bottom. The absolute position deserves more caution. With 429 votes the interval widens to ±26, and the board reports a rank spread of 1 to 3. "Top three on video editing preference" holds up today. "The best video editing model" does not.
Why Artificial Analysis shows no Seedance 2.5 score
Artificial Analysis Text to Video Leaderboard (With Audio), captured 2026-09-17. Source: Artificial Analysis Text to Video Leaderboard
Artificial Analysis runs its own Video Arena with its own voter pool and its own Elo scale, and it adds models on its own schedule. On 2026-09-17 the ByteDance Seed entries on its Text-to-Video board were Dreamina Seedance 2.0 720p at 5th with 1210 Elo across 20,345 samples, plus Seedance 1.5 pro sitting at exactly 1000, the board's anchor value. The Image-to-Video board looks the same, with Seedance 2.0 720p at 5th on 1174 Elo. On the Video Editing board Seedance 2.0 720p sits 6th at 1035, just ahead of Runway's Aleph 2.0 at 1011 and Kling 3.0 Omni 1080p (Pro) at 1000.
Two things follow. If a comparison article quotes an Artificial Analysis score for Seedance 2.5, that score is not on the board it claims to come from. And the scales themselves do not line up: Seedance 2.0 reads 1210 on Artificial Analysis and 1479 on Arena.ai. The two cannot be averaged, subtracted, or dropped into the same column as though they shared a unit.
VBench, the third source people reach for, does not help here. Its public leaderboard has never carried a Seedance model, and its most recent VBench-team-certified entries are dated 2025. As an automated breakdown of 16 quality dimensions for the open-source models it does cover, VBench remains useful. It cannot answer a 2026 Seedance versus Veo question.
What each leaderboard actually measures
| Source | Method | What it can tell you | What it cannot |
|---|---|---|---|
| Arena.ai Video Arena | Blind pairwise human votes on user-supplied prompts, scored Elo-style with confidence intervals and rank spreads | Which model people currently prefer on short, generic prompts | Anything about long-form work, specific briefs, or non-default settings |
| Artificial Analysis Video Arena | Blind pairwise human votes, separate voter pool and prompt mix, split into With Audio and No Audio boards | Preference ordering inside its own ecosystem, plus a cost-per-minute column | Anything about Seedance 2.5 today, and its scores never transfer to Arena.ai's scale |
| VBench | Automated scoring across 16 defined quality dimensions, submission-based with tiered certification | Reproducible dimension-level detail for the models submitted to it | Any Seedance version, and anything released after its last refresh |
The practical rule: use these boards to place a model in a tier, not to rank it against a rival two positions away. Any gap smaller than the sum of the two confidence intervals is noise, and on the Text-to-Video board that threshold sits at roughly 20 points, wider than the distance separating four of the top ten models.
What the boards do not cover
The three capabilities Seedance 2.5 was built around are the three no public arena currently votes on. Single-pass 30-second clips, up to 50 reference inputs in one generation, and timestamp-level edits all sit outside the short, default-setting comparisons the arenas run. A model that ties its predecessor on a six-second prompt can still behave very differently across a 30-second narrative.
That is why scenario-level comparisons stay worth reading next to the score tables. Our head-to-head on Seedance 2.5 versus Kling 2.5 for character and motion consistency and the breakdown of Seedance 2.5 versus Veo 3.1 across prompt adherence and audio both work on longer briefed shots rather than single-prompt votes, which is where the differences the arenas flatten tend to appear.
Cost is the other axis the preference boards mostly skip. Artificial Analysis publishes a cost-per-minute column beside its Elo values, Arena.ai does not, and neither adjusts a score for price. The AI video tool price comparison sets the per-second economics side by side, and current per-plan rates live on the SeedVideo AI pricing page.
How to check these numbers yourself
- Open the Arena.ai board for the category you care about, whether that is Text-to-Video, Image-to-Video, or Video Edit, and note the date printed under the title. Each category updates on its own schedule.
- Read the score with its ± interval and its rank spread. Never the rank on its own.
- Check the vote count. Below roughly 5,000 votes, treat the position as provisional.
- Open the Artificial Analysis board separately and keep its scores in their own column.
- Re-check monthly. Seedance 2.5's Video Edit position has already shifted relative to Wan 3.0 since the model arrived.
FAQ
What is Seedance 2.5's benchmark score?
On Arena.ai's Video Arena, verified 2026-09-17: 1482 ±12 in Text-to-Video (rank 6 of 48), 1475 ±8 in Image-to-Video (rank 4 of 48), and 1410 ±26 in Video Edit (rank 2 of 10). These are blind human-preference scores on that board's scale only. Seedance 2.5 has no Artificial Analysis score and no VBench score.
Is Seedance 2.5 better than Veo 3.1 according to benchmarks?
On the Arena.ai boards it scores higher than every Veo entry: 1482 against veo-3.1-audio's 1364 in Text-to-Video, and 1475 against 1398 in Image-to-Video. That reflects blind preference on short generic prompts with audio enabled. It says nothing about prompt adherence, resolution ceiling, or long-form consistency, and no Veo model is entered in the Video Edit category at all.
Why is Seedance 2.5 missing from the Artificial Analysis leaderboard?
Artificial Analysis adds models to its Video Arena on its own schedule, and as of 2026-09-17 it has not added Seedance 2.5 to any of its three video boards. Its newest ByteDance Seed entry remains Dreamina Seedance 2.0 720p. Any Artificial Analysis figure attributed to Seedance 2.5 did not come from the live board.
Does Seedance 2.5 outrank Kling and Runway?
Yes, across all three Arena.ai video boards as of 2026-09-17. The strongest Kling entries are kling-2.5-turbo-1080p at 1219 in Text-to-Video, kling-v3-pro at 1355 in Image-to-Video, and kling-o3-pro at 1255 in Video Edit. The strongest Runway entries are runway-gen-4.5 at 1224, runway-gen4-turbo at 1052, and runway-gen4-aleph at 1182 in the same three categories. Runway's flagship Gen-4.5 is entered only in Text-to-Video.
How much better is Seedance 2.5 than Seedance 2.0 on benchmarks?
Less than the headlines imply, on two boards out of three. Text-to-Video is 1482 against 1479, a 3-point gap measured against a ±12 interval. Image-to-Video is 1475 against 1474, a 1-point gap against ±8. Both are statistical ties. Video Edit is the exception at 1410 against 1365, a 45-point gap that holds. The things that separate the two versions in practice, such as 30-second single-pass generation and 50-reference input, fall outside what these boards test.
How often do these scores change?
Each Arena.ai category refreshes independently. On 2026-09-17 the three video boards carried dates of Sep 4, Sep 14, and Aug 26, 2026. Models with low vote counts move most. We re-check this page monthly and update the Last verified line on each pass.
Where to go from here
If you came for a single number, the honest one is this: Seedance 2.5 currently sits in the top tier of one public video arena, is absent from the other, and has a measured lead over Seedance 2.0 only in video editing. Any score quoted without a board name, a date, and a vote count is not usable.
For the capability list the arenas never touch, covering clip length, reference limits, and editing controls, start from the Seedance 2.5 overview, then run your own brief through SeedVideo AI and judge the output against the shot you actually need.
