Artificial Analysis re-based its image-to-video leaderboard

A leaderboard position is a snapshot of a methodology, not a property of a model. Artificial Analysis re-based its image-to-video ranking this month, the old Elo scores stopped carrying over, and at least one vendor's "we are No. 2" claim now survives on only one of the board's two audio toggles. If you put a rank in a pitch deck, a price page or a launch post last month, the number you copied may not exist on the board any more.

The image-to-video board is the benchmark most vendors quote at launch, and two things moved on it this month: the version, and the models measured on it.

What Artificial Analysis changed

The board's own header now reads AA-Video-I2V v1.1, and it "Ranks video models on animating a source image into video with audio, by Elo from pairwise human preference votes." A banner on the same page says v2.1 is coming: "AA-Video-I2V v2.1 is coming soon, with a new prompt set aligned with the AA-Video-T2V v2.1 use case and capability taxonomy, 1080p videos on the silent board, and a more comprehensive model selection."

The methodology page explains why a single rank cannot be lifted out of context. It lists six benchmarks, three with audio and three silent, and says: "Each benchmark has its own Elo pool in the Video Arena, so silent and audio-enabled outputs are not compared directly, and editing is not compared with generation." Every board is anchored so that Kling 3.0 1080p (Pro) sits at exactly 1000.

That structure matters: a rank on the with-audio board is a position inside one pool, and you cannot move it to another board without saying which pool it came from.

Why the old numbers do not carry over

The v1.1 change is not a relabel. The methodology page says v1.1 moves image-to-video with audio to a native resolution standard: each model is served at its highest native resolution up to 1080p at 10 seconds, and models that cannot generate 10 seconds are served at their maximum length. Models whose earlier clips missed that standard "were regenerated at it, and votes on their earlier clips were removed." Only votes cast between models on the v1.1 leaderboard count toward its Elo.

A second change came in October 2026 to the statistics rather than the models. The Elo calculation, the confidence intervals and the Range column "follow an updated, more rigorous statistical method," and the confidence intervals "now include the uncertainty in every other model's rating." The range you see beside a model is no longer just that model's own error bar.

Together those two changes mean the previous ranking and this one are not comparable. RCTV's 19 October roundup put it plainly: "Elo values do not carry over from the earlier version, so the old top five with Vidu Q4 Preview at third (1179) cannot be compared with this one." Vidu Q4 Preview was third on the board a vendor would have screenshotted last month, and the 1179 no longer sits on any scale you can point a reader at.

What the with-audio board says now

Read on 20 October 2026, on the AA-Video-I2V v1.1 board with the With Audio toggle selected, the top of the table is MiniMax H3 (768p) first at 1140, then FLUX 3 Max (720p) and Dreamina Seedance 2.5 tied at 1138, then Gemini Omni Flash 1.1 at 1132, then Wan 3.0 at 1127. MiniMax H3 Max is sixth at 1125, FLUX 3 is eighth at 1086, Vidu Q3 Pro is 17th at 1016 and Vidu Q3 Turbo is 21st at 967.

Those top five sit inside 13 points, from 1140 down to 1127, against confidence intervals of ±15 to ±17. RCTV, reading the same board two days earlier, described the spread as "within 15 points and every interval is ±16 or ±17, so the board has no clear leader." Either way the board does not separate the leaders: a "No. 1 on Artificial Analysis" line from this board is a claim inside a 13-point band that overlaps with four rivals.

The board also moved between reads. RCTV read it on 18 October and reported FLUX 3 seventh at 1086. By my read on 20 October, FLUX 3 Max (720p) had entered at second with 1138 and pushed FLUX 3 to eighth at 1086. The page itself notes three models added in the last 30 days: FLUX 3 Max (720p), MiniMax H3 (768p) and Dreamina Seedance 2.5.

The vendor claims that moved with the re-base

Video Rebirth is the clearest case of a claim that did not travel with the board. Its own post on X says: "BACH 1.0 Pro is No. 3 on the @ArtificialAnlys Image to Video leaderboard. Four months ago, BACH 1.0 Preview entered the same leaderboard at No. 5. What happened in between: we let the model teach itself." A May 2026 company press release described BACH 1.0 Preview as "ranked #6 in the world on Artificial Analysis upon debut." TipRanks also wrote up the same claims from Video Rebirth's LinkedIn post.

Those are the vendor's claims, and they do not agree with each other: the same model is No. 6 on debut in the press release and No. 5 on entry in the later post. Video Rebirth does not reconcile them.

RCTV reported that Video Rebirth launched BACH 1.0 Pro commercially on 15 October as "Currently No. 2" on the image-to-video leaderboard. On the no-audio board, BACH 1.0 Pro sits second at 1362, level with Wan 3.0 at 1362 and six points behind Gemini Omni Flash at 1368. On the with-audio board, there is no BACH row at all. So "No. 2" is true only if you name the silent board, and "No. 3" is true on neither board I read. RCTV noted the same split, and Video Rebirth's own site names the toggle and lists $0.12 a second at 1080p.

Vidu Q4 is on neither board, and Vidu's pages disagree on its name

There is no Vidu Q4 Preview row on the with-audio board I read, and no Vidu Q4 row of any kind on the no-audio board. The only Vidu models present are Vidu Q3 Pro, 17th at 1016, and Vidu Q3 Turbo, 21st at 967, both on the with-audio board. RCTV reported the same absence more broadly: "Vidu Q4 Preview is missing from every board and toggle we read, along with HiDream-O1-Video-1.0. The page has filters that hide research previews and derived models, so a default setting could explain it. AA has not said."

Artificial Analysis has not explained the absence, and its board carries "Hide research previews" and "Hide derived models" filters that could account for a preview model dropping out of a default view. Do not report a cause the site has not given.

Vidu's own pages disagree about the name. Its API model map lists "ViduQ4 Preview" under the model ID viduq4-preview, and the pricing page bills that same ID from 9 credits per second at 540p to 24 at 1080p, 38 at 2K and 78 at 4K. The marketing page at vidu.com/vidu-q4 drops the qualifier and calls it "Vidu Q4 AI Video Model with Native Audio," listing 4K maximum resolution and up to 16 seconds. A reader citing "Vidu Q4" and a developer calling viduq4-preview are naming the same model from two pages that write its name differently.

How to cite a rank so it stays true

On these boards the version, the toggle and the date are not context around the claim, they are the claim. RCTV's own line is the rule to copy: "Cite a rank with the board version, the audio toggle and the date, or don't cite it."

Before you repeat any rank from this board, check four things. First, the version string at the top of the page: v1.1 is current, v2.1 is announced and not live, and a number from v1.0 is not a number you can put beside a v1.1 figure. Second, which toggle the figure came from: with-audio and no-audio are separate Elo pools, and the same company can sit second on one and be absent from the other. Third, the read date, because the roster changes: three models joined in the last 30 days alone. Fourth, the interval, because a two-point lead inside a ±15 band is a tie, not a lead.

The vendor's claim is the last thing to check, not the first. BACH 1.0 Pro's "No. 2," its "No. 3" and its old "#6" each describe a real board state, and none of them tells you which version, toggle or week it came from. When a vendor's copy and the live board disagree, the board is what changed, and the version label and toggle are how you say so.

Both boards are free to read at artificialanalysis.ai/video/leaderboard/image-to-video, where the With Audio and No Audio toggle sits above the table beside the version label; copy the label, the toggle and the date with any number you take from it.