Open-weights AI video model reaches top 3 in 2026 ranking

MarkTechPost has published an updated ranking of the best AI video generation models for 2026, and it shows a leaderboard that looks markedly different from a year earlier. According to the outlet, an open-weights model now sits in the top three, native audio is treated as a standard feature rather than a bonus, and typical generated clips run between 15 and 30 seconds — shifts that suggest how quickly the category is maturing.

How has the AI video generation leaderboard changed in a year?

The clearest takeaway from MarkTechPost's ranking, dated November 13, 2026, is how much has moved in twelve months. The outlet frames its list around a single quality signal — blind human-preference Elo drawn from the Artificial Analysis Video Arena — and then pairs that score with each model's API price per second. That combination is meant to answer two questions at once: which models people actually prefer when they cannot see the brand behind them, and what each of those models costs to run.

Three changes stand out. Native audio, meaning sound generated together with the picture, is described as table stakes rather than a differentiator. Clip length has settled into a 15-to-30-second range. And, notably, an open-weights model now sits inside the top three — a strong showing for a downloadable model in a category where the highest-profile tools have typically been closed and reachable only through an API.

What is blind human-preference Elo, and where does it come from?

Elo is a rating system borrowed from chess, where players gain or lose points based on the outcome of head-to-head matches. Applied to AI models, it works by showing people pairs of outputs and asking which they prefer; the winner's rating rises and the loser's falls. The blind part matters: raters do not know which model produced which clip, which is designed to remove brand loyalty and reputation from the judgment.

According to MarkTechPost, the Elo figures in this ranking come from the Artificial Analysis Video Arena, a public comparison platform. Using one consistent quality signal across every model is what makes the leaderboard legible — it places very different systems on the same scale, so a preference score can be read next to a price without mixing unrelated measures.

Why is native audio now considered table stakes?

Native audio means the model produces synchronized sound — dialogue, ambient noise, effects — as part of generating the video, rather than leaving creators to add a soundtrack or dubbing in a separate step. MarkTechPost reports that this capability has moved from a selling point to a baseline expectation over the past year.

In practice, that changes the workflow. When audio arrives attached to the footage, the manual work of matching voice to lip movement or layering in effects shrinks, and a single prompt can yield a clip that is closer to finished. It also raises the floor for the whole category: a model that produces only silent video now looks dated next to ones that generate picture and sound together.

What does an open-weights model in the top 3 mean?

An open-weights model is one whose trained parameters are published and can be downloaded, run on your own hardware, or fine-tuned, as opposed to a closed model that is only reachable through a vendor's API. According to MarkTechPost, one such model has reached the top three of the 2026 ranking.

That matters for a few reasons. Open weights give teams control over cost, since running a model yourself can be cheaper than paying per generation at scale. They also allow customization and keep sensitive material on infrastructure you control. And a strong open-weights entry near the top signals that the gap between the best closed services and downloadable models has narrowed enough to matter for real projects.

Why does price per second matter alongside quality?

MarkTechPost deliberately pairs its Elo score with each model's API price per second, and that pairing reflects how video generation is actually billed. Because you pay for the length of what you generate, price per second is the practical unit of cost — a model that looks better but costs several times more may not be the right choice for high-volume work.

Reading quality and price together is more useful than reading either alone. A model near the top of the Elo table but expensive per second suits hero shots and final renders, while a cheaper model further down might be the sensible default for drafts and experiments. The ranking's structure encourages that trade-off thinking rather than a single best pick.

How can you try these video models in Grafy?

If you want to test the differences the ranking describes, Grafy's Creator generates video from a prompt or a photo and gives you many cloud models side by side, so you can compare how different systems handle the same brief, including whether they return native audio. For turning short clips into something longer, Film Studio helps plan and assemble AI documentaries and short films, which is useful once you are stitching together several 15-to-30-second generations. And when you need to trim, cut or finish the footage, Video Edit handles editing and rendering. Together they cover the loop from a single prompt to a finished piece, without forcing you to commit to one model up front.

Frequently asked questions

Which quality signal does MarkTechPost use for its 2026 AI video ranking?

According to MarkTechPost, the ranking uses one quality signal: blind human-preference Elo drawn from the Artificial Analysis Video Arena. In blind comparisons, raters judge pairs of clips without knowing which model made each one. The outlet then pairs that Elo score with each model's API price per second, so quality and cost can be read together.

Does the 2026 ranking include an open-weights model in its top 3?

Yes. According to MarkTechPost, an open-weights model sits in the top three of the 2026 ranking. Open weights mean the model's parameters are published and can be downloaded, self-hosted or fine-tuned, rather than being reachable only through a vendor's API. Its presence near the top signals that downloadable models have closed much of the gap with closed services.

How long do the AI-generated clips in the ranking typically run?

According to MarkTechPost, generated clips in the 2026 ranking typically run between 15 and 30 seconds. That range has become the common length for a single generation. For longer pieces, creators generally assemble several of these short clips and edit them together, which is where tools for planning, editing and rendering enter the workflow.

---

Source: MarkTechPost — Best AI Video Generation Models in 2026: Elo, Price per Second, Open Weights