AI Video in 2026: Wan 3.0 Leads at 1175 Elo, MiniMax H3 Is the Open-Weights Base

By Marcus Bennett

Alibaba's Wan 3.0 tops the Artificial Analysis text-to-video leaderboard with audio at 1175 Elo, as of November 13, 2026, and it costs $0.20 per second through the API. The 2026 video generation market looks nothing like it did twelve months ago: native audio is now table stakes, clips run 15–30 seconds, and an open-weights model — MiniMax H3 — sits at number three. At least four of the top fourteen models are fine-tunes built on its weights.

The ranking comes from blind human-preference voting in the Artificial Analysis Video Arena, where validated evaluators compare two clips generated from the same prompt. Kuaishou's Kling 3.0 1080p (Pro) is fixed at 1000 Elo as the anchor on both the text-to-video (AA-Video-T2V v2.1) and image-to-video (AA-Video-I2V v2.1) boards; any score below 1000 means voters preferred Kling 3.0 more often. Most top-10 scores carry a ±13 to ±15 Elo confidence interval, so models within roughly 25 points are statistically close — ranks 3 through 8 function as a cluster rather than a strict order. Prices are Artificial Analysis per-minute API costs, normalized to 1080p where offered, divided by 60.

Who leads each category?

• Text-to-video: Wan 3.0 at 1175 Elo ($0.20/sec)
• Image-to-video: ByteDance Dreamina Seedance 2.5 at 1161 Elo ($0.569/sec)
• Open weights: MiniMax H3 at 1161 Elo ($0.08/sec)
• Price efficiency: HeyGen Video 1 at 1152 Elo ($0.04/sec)
• Speed: fal's FLUX 3 Max, ~12 seconds for a 10-second 720p clip ($0.145/sec)

The top ten, by the numbers

1. Wan 3.0 (Alibaba) — 1175 T2V Elo, 1108 I2V Elo, $0.20/sec, closed weights. Launched August 24, 2026, with 7,651 votes behind its score, Wan 3.0 generates 30-second 1080p clips from text, PDFs and slide decks. The document-to-video input is the differentiator; the trade-offs are no downloadable weights — Wan 2.2 from July 2025 remains Alibaba's last open video flagship — and a drop to 11th place on image-to-video.

2. Utopai X (Utopai Studios) — 1170 T2V, 1151 I2V, no public API, closed weights. The Mountain View film studio post-trained MiniMax H3 and beat the base model. It is the only model in the top two on both boards. Access runs only through Utopai's PAI platform at 93 credits per second, with plans starting at $15 for 1,000 credits. Developers cannot call it from a pipeline.

3. MiniMax H3 (MiniMax) — 1161 T2V, 1133 I2V, $0.08/sec, open weights under the MiniMax H3 Community License. The API launched July 31, 2026; weights followed on August 3. It is a 33B dense single-stream Transformer using Qwen3-VL-32B as its encoder, outputting video with native 32 kHz stereo audio.

Dreamina Seedance 2.5 (ByteDance Seed) — 1155 T2V, 1161 I2V (first place), $0.569/sec on the normalized basis, closed. Rolling out since July 31, 2026 through Jimeng AI and Doubao Pro, it supports up to 30-second generations with up to 30 image, 10 video and 10 audio references. Pricing needs care: Artificial Analysis lists $34.12 per minute, the highest on the board, while BytePlus lists $1.156 for a 5-second 720p clip without video input — about $0.23 per second.

Gemini Omni Flash 1.2 (Google) — 1154 T2V, 1086 I2V, $0.104/sec on the normalized basis ($0.07/sec at 720p per Google), closed. Generally available since November 5, 2026, it added 3–15 second clips and a 30% price cut over version 1.1, jumping 27 Elo on text-to-video. Requests to 1.1 route to 1.2 from December 3, 2026.

FLUX 3 Max (fal × Black Forest Labs) — 1153 T2V, 1068 I2V, $0.145/sec, closed. Launched October 19, 2026, it renders a 10-second 720p clip in about 12 seconds, according to fal; Seedance 2.5 took 277.80 seconds in fal's own comparison. It shares an API with fal's H3 Max — switching takes a single model-ID change.

HeyGen Video 1 (HeyGen) — 1152 T2V, 1124 I2V, $0.04/sec, closed. A post-train of MiniMax H3 positioned as a general scene generator rather than an avatar engine, it leads the silent (audio-off) text-to-video board at 1146 Elo and costs half of base H3 for a similar score. The API model ID is heygen-video-1.

MiniMax H3 Max (fal) — 1149 T2V, 1115 I2V, $0.16/sec, API only. fal Research post-trained the open MiniMax H3 base, focused on prompt adherence and visual quality, and claims a 5-second clip renders in about 3 seconds. It briefly led image-to-video after its August 2026 launch before newer post-trains passed it.

SkyReels V4 (Skywork AI) — 1137 T2V, 1125 I2V, $0.35/sec, closed. Dating from March 2026, it still ranks fifth on image-to-video but is hard to justify on text-to-video: Gemini Omni Flash 1.2 scores higher for less than a third of the price.

FLUX 3 (Black Forest Labs) — 1137 T2V, 1071 I2V, $0.29/sec, closed today. It is BFL's first video model, producing 20-second clips with native audio; one backbone also handles robot action prediction. BFL says it will release open-weight versions later this year; until FLUX 3 Dev ships, treat it as API-only.

What are the open-weights options?

MiniMax H3's Hugging Face weights carry real restrictions. The license blocks local deployment in the US, EU, UK and South Korea; only H3-Base is open with a 768p ceiling (2K upscaling stays API-only); and the repo holds two task-specific checkpoints totaling 288 GB. Pika H3 Flex, shipped October 29, 2026 under the same license, cuts H3 to 6 sampling steps but local inference still needs an H100 80GB and about 215 GB of downloads. Video Rebirth's HyperFlow H3 Flash compresses H3 to 13.7B, running from a 9.1 GB GGUF on a 24 GB GPU; its sibling, Hyperflow H3, ranks seventh on image-to-video at 1120 Elo.

Lightricks released LTX-2.5, a 22B model with native multishot generation and joint audio, on August 11, 2026. Under the LTX-2.x Community License, entities with $10M+ annual revenue need a paid license. It trails the H3 family by roughly 200 Elo (LTX-2.5 Fast scores 943, LTX-2.5 Pro 938) but gates by revenue rather than territory — a practical default for US and EU teams. For permissive licensing, Wan 2.2 under Apache 2.0 and Tencent's HunyuanVideo 1.5 (8.3B, aimed at single consumer GPUs) remain the standard picks, though neither appears on the current Artificial Analysis boards.

Why does price diverge from quality?

Quality and price no longer track each other. The three cheapest top-tier models — HeyGen Video 1 ($0.040/sec, 1152 Elo), Pika H3 Flex ($0.035/sec, 1124 Elo) and MiniMax H3 ($0.080/sec, 1161 Elo) — are all H3 derivatives or H3 itself. Measured as Elo above the 1000 anchor per dollar per second, HeyGen Video 1 leads at 3,800, ahead of Pika H3 Flex at 3,543 and H3 at 2,012; Seedance 2.5 sits at just 273. Wan 3.0 buys roughly 23 Elo over HeyGen Video 1 at five times the price per second — whether that matters depends on how many takes a production discards.

Where do Veo, Kling 4.0, Sora and Runway stand?

Google's Veo 3.1 scores 964 on text-to-video, below the Kling 3.0 anchor; Veo 3.1 Lite matches base H3 on price at $0.08/sec but trails it by 213 Elo. Kling 4.0 became generally available on klingai.com with 30-second native clips, but Kuaishou had not announced API availability or pricing at last check, and it does not yet appear on the boards. OpenAI's Sora 2 and Runway's Gen-4.5 are absent from the current with-audio leaderboards; the ranking's compilers excluded them rather than mix in non-comparable scores.

The practical shortcuts: Wan 3.0 for the highest text-to-video quality with API access; Seedance 2.5 for image-to-video and heavy references; HeyGen Video 1 or Gemini Omni Flash 1.2 for quality per dollar; FLUX 3 Max or H3 Max on fal for iteration speed; MiniMax H3 or Pika H3 Flex for self-hosting outside the US, EU, UK and South Korea; LTX-2.5 or Wan 2.2 for self-hosting within them. With H3 derivatives crowding the value tier and Kling 4.0 still unlisted, the December 3 routing cutover to Gemini Omni Flash 1.2 is the next concrete date that could reshuffle the top five.