news

HeyGen Video 1 leads Artificial Analysis's silent text-to-video board by 3 Elo, and two AA pages disagree on its score

The model is listed first on AA-Video-T2V-Silent at Elo 1143, ten points clear of Wan 3.0, while the same board's own published margin on that score is plus or minus 15.

By Priya Shenoy · 30 October 2026

Artificial Analysis currently lists HeyGen Video 1 at the top of its silent text-to-video Elo board, ahead of MiniMax H3 Max, FLUX 3 Max and Wan 3.0. The four models are separated by a total of ten Elo points, which is smaller than the margin Artificial Analysis publishes next to the individual scores, and two of Artificial Analysis's own pages do not agree on what HeyGen Video 1 scored or where it ranks.

An avatar video company at the top of a general text-to-video ranking

Artificial Analysis describes AA-Video-T2V v2.1 as ranking video models on generating video with audio from a text prompt, by Elo from pairwise human preference votes. The embeddable version of the same board adds that the quality evaluation is based on blind preferences from what the company calls its validated private evaluators, and that the image-to-video board without audio also folds in public Video Arena votes cast before 1 January 2026. HeyGen Video 1 is one of five models the board says were added in the last 30 days, alongside FLUX 3 Max and MiniMax H3 Max.

On the with-audio board, Wan 3.0 leads at Elo 1174, Utopai X follows at 1168, MiniMax H3 at 768p sits at 1158 and HeyGen Video 1 at 1157. On the silent board the order flips: HeyGen Video 1 at 1143, MiniMax H3 Max at 1140, FLUX 3 Max at 1138 and Wan 3.0 at 1133.

The gap between first and fourth is smaller than the stated uncertainty

HeyGen Video 1's with-audio score is published as 1157 plus or minus 15, with a 95% interval running from 1142 to 1172, and Artificial Analysis's own rank range column places the model anywhere from second to seventh. The single Elo point that separates it from MiniMax H3 at 768p is not a result in any statistical sense, and the three-point lead on the silent board sits well inside the same kind of interval. The ordinal position is the headline, but the band is the finding.

The same model, two ranks, two sample counts

The main leaderboard page and the embeddable leaderboard page do not match. HeyGen Video 1 appears on one as rank 4, Elo 1157 plus or minus 15, on 5,714 samples, and on the other as rank 3, Elo 1160 plus or minus 15, on 5,520 samples. FLUX 3 Max appears on one as Elo 1156 plus or minus 18 on 2,974 samples, and on the other as 1155 plus or minus 15 on 5,479 samples.

Leaderboards refresh, so two surfaces showing different snapshots is ordinary. What is harder to explain is the direction: the main page carries more HeyGen samples than the embed, while the embed carries nearly twice as many FLUX 3 Max samples as the main page. One page is not simply a stale copy of the other, which means a reader cannot tell which set of numbers is current without being told.

The price column is the hardest number to compare

Artificial Analysis states that its pricing column reflects the cost to generate one minute of video on the model creator's API at the settings it benchmarks, 1080p where offered. FLUX 3 Max is listed at $4.35 per minute on the main text-to-video board and on the image-to-video board, and at $2.55 per minute on the embeddable board, where the row is labelled FLUX 3 Max (720p) rather than FLUX 3 Max. HeyGen Video 1 is listed at $2.40 per minute on both. On image-to-video, FLUX 3 Max sits at rank 7 with Elo 1064 plus or minus 17 on 5,242 samples.

Two rows with near-identical names, different resolution labels and a price difference of 70% are not interchangeable inputs to a cost-per-quality comparison. For anyone picking a model off these boards, the usable reading is the rank range and the interval, not the ordinal.

Sources: Artificial Analysis AA-Video-T2V v2.1, Artificial Analysis text-to-video embed leaderboard, Artificial Analysis AA-Video-I2V v2.1.

Priya Shenoy: Tracks what AI video actually costs across the platforms that resell access to the same handful of models. Treats a pricing page as a claim, not a fact, until someone checks it.