The state of AI video, November 2026

Two problems make this harder than it should be. There is exactly one independent test of AI video quality, and it was reset over the summer, so most of the numbers in circulation cannot be compared with each other. And roughly four-fifths of what a search returns on this subject is content-farm output carrying fabricated specifications and invented release dates, including for models that do not exist. What follows is the map as of November 2026, with the tested claims kept separate from the marketed ones.

The only independent test

Artificial Analysis runs blind human preference voting, which is the one measure here not supplied by a vendor. The late October 2026 category results: Wan 3.0 leads text-to-video overall. Seedance 2.5 leads image-to-video overall, and also photorealism, camera control, human anatomy, retail and ecommerce, and cartoon and anime. MiniMax H3 leads on-screen text rendering. MiniMax H3 Max leads spatio-temporal consistency. Seedance 2.0, the previous generation, still leads audio sync and lighting and materials. Category breakdown here.

The caveat matters more than the table. The board was reset in summer 2026, so any score from before August is not comparable to a current one. That retires a figure you will still be shown: Runway's Gen-4.5 launched on 1 December 2025 at number one on text-to-video with 1247 Elo. True when it was published. Not a claim about today.

Google

Gemini Omni 1.1 Flash went generally available on 27 August 2026, model ID gemini-omni-1.1-flash. Clips of three to ten seconds, scene extension in ten-second increments to a cumulative forty, and output at 360p, 720p, 1080p or 4K. The resolution line needs reading carefully: 1080p and 4K are upscaled, not native. Release detail. Veo 3.1, from 15 October 2025, remains the enterprise model, and Veo 3.1 Lite shipped on 31 March 2026 at under half the cost of Veo 3.1 Fast.

There is no Veo 4. Every page offering its specifications, benchmarks or release date is invented. If a vendor, an agency or a search result quotes Veo 4 at you, you have learned something about the source rather than about Google.

Runway

Gen-4.5 is the headline model, but Aleph 2.0 and Edit Studio, from 21 May 2026, are the more consequential release for commercial work. In-context editing of existing footage up to thirty seconds at 1080p. Look transfer from a reference frame. One edit propagating across multiple shots and through scene changes. And a still preview before you spend a render, which changes the cost of being wrong.

Free

125 credits

Standard

625

Pro

2,250

Max

9,500

Pricing is per-credit: Free at 125 credits, Standard $15 a month for 625, Pro $35 for 2,250, Max $95 for 9,500. One detail decides it for anyone finishing in a grade. Max is the only tier with HDR, ProRes and image-sequence output. If the film goes to an online, the cheaper tiers are not a cheaper version of the same thing.

ByteDance Seedance 2.5

Released July 2026 and the strongest single model on the independent board for image-to-video. Thirty-second clips. Native 4K at 10-bit, explicitly not upscaled, which currently separates it from Google. Up to fifty multimodal references, against twelve in the previous version, and the reference types now include audio, 3D white models and style references. Audio generated in the same latent space as the picture. A 3D white-box previz pass. Long-video beta out to three minutes. Specifications here.

The claim to discount is the 20% improvement in prompt adherence. That is ByteDance's own number with no independent test behind it. Treat it as marketing until the board measures it.

Alibaba Wan 3.0

Public beta from 6 August 2026. Up to thirty seconds at 480p, 720p or 1080p, with no 4K. Audio in the same pass as picture. Document-to-video from doc, xls, ppt, pdf and markdown files, plus URLs. API only, closed weights. Model page.

“If a vendor, an agency or a search result quotes Veo 4 at you, you have learned something about the source rather than about Google.”

Sit with the combination for a second, because it contradicts how this category is usually sold. The model that leads text-to-video on the only independent test tops out at 1080p. Resolution and preference are not the same axis.

MiniMax H3

Hailuo 3.0, announced 31 July 2026. Up to 2K, meaning 1,440px on the short edge, at 24fps. Four to fifteen seconds. Native stereo audio. Reference-to-video accepting nine images, three video clips and three audio clips. Pay-as-you-go at $0.13 a second at 2K. Full specification. Open weights, which is the part with consequences further down this list.

Kling 4.0

Announced 28 September 2026 and live on resellers by 20 October. Three to thirty seconds at 720p or 1080p, up to ten keyframes, fifteen references, two-channel stereo audio, 21:9, on-screen text in nine languages and prompts up to 8,000 tokens. Reseller pricing works out around $11.03 for thirty seconds at 1080p. Spec sheet, reseller detail.

4K and 10-bit HDR are listed as coming soon. They have not shipped. Do not plan a 4K delivery around them.

Luma Ray3.2

From 9 June 2026. Up to twenty seconds at 1080p, up to sixteen keyframes, and performance tracking of skeletal posture, gesture and facial expression for as many as eight faces in frame. Native HDR with 16-bit EXR export, which is the cleanest handoff into a finishing pipeline on this list. Dream Machine is deprecated. Announcement.

Pika, Adobe, Midjourney, Sora

Pika H3 Flex, from 29 October 2026, is a post-trained MiniMax H3: up to twelve references, five separate audio stems, finishing to 4K, and fifteen seconds of synced video with stereo audio in six denoising steps, with open weights. Pika's blog.

Adobe Firefly hosts Veo 3.1, Luma Ray3.14, Runway Gen-4.5 and Kling 3.0 in one interface. What it offers is clearance and a single workflow, not capability. It has no frontier model of its own, and buying it for quality is a category error.

Midjourney: we could not verify its current video version or any of its specifications from a non-SEO source, so we are not going to state any. Sora is discontinued. The app closed on 26 April 2026 and the API was switched off on 24 September. Treat it as unavailable, including in any pitch deck still quoting it.

Cost, which is where the decision gets made

Cost per minute, from the same independent board: Seedance 2.5 at $34.12, Seedance 2.0 at $22.45, MiniMax H3 at 768p for $4.80, H3 Max at $5.76. Seedance 2.5 is roughly seven times the cost of H3 per minute of output.

That ratio, not the rankings, is the real pipeline decision. A thirty-second hero that finishes in a 4K grade earns the expensive model. Forty social cutdowns do not, and spending Seedance money on them is how an AI production budget ends up matching a conventional one with none of the control to show for it. We quote the two separately on the same job, because they are two different purchases.