Runway’s Gen-4.5 Sets New Bar for AI Video Generation

Big tech wasn’t expected to come in second in this race. For the past year, it has generally been presumed that the most believable AI videos would be created by companies that maintain the largest data centers and the largest research facilities. Contrarily, the current top model on the leaderboard is owned by Runway, a video-centered startup located in New York.

Runway launched Gen-4.5 on December 1 as its latest text-to-video model. According to the company, Gen-4.5 can generate video that resembles real-world camera captures better than previous systems, demonstrating better control and quality. In addition, Runway asserts that Gen-4.5 is currently at the top of the popular leaderboard for text-to-video models in independent testing, outperforming competitors developed in larger labs.

What Gen-4.5 actually improves

Runway’s pitch for Gen-4.5 is simple: videos should look better and follow directions better.

In terms of quality, the addition of realistic weight and speed for moving objects, and more realistic fluid movement that accurately captures gravity and impact, is key. In early demos, characters move in more realistic ways, and scenes involving several objects such as cars, animals, and light stay fixed for several seconds of animation rather than degrading shortly after the first frames. According to The Verge, it is “cinematic and highly realistic” and has dramatically increased the accuracy of scenes that correspond to the textual description.

Gen-4.5 allocates considerable computing power to detailed work too. Runway emphasizes the importance of keeping hair, fabric, and reflections steady from one frame to the next. They also note that stylistic decisions like documentary-style video, stop-motion animation, or highly graded “movie” looks, now hold together for the entire video, rather than looking good for the first moment and then fading.

These visual improvements continue the quick turnaround cycle that Runway has maintained for the past 18 months. Gen-3 Alpha,launched in 2024, increased clip length and expressiveness of characters and ended up on Time Magazine’s “Best Inventions” list. Gen-4, launched earlier this year, optimized for keeping characters and locations in sync for multi-shot workflows, so that small stories and ads can be created, not just hero clips.

 Gen-4.5 carries on this tradition: it keeps the multi-shot workflows, improves the realism of physics and details, and comes in the same modes: text-to-video, image-to-video, and video-to-video within the Runway app and API.

Benchmarks and the “Number One” Claim

Runway's “best video model” claim is based on Artificial Analysis’ Video Arena, which has emerged as one of the very few public scoreboards in this domain. Two side-by-side video clips are showcased, the viewer selects which one they prefer, and the system converts those judgments into “Elo ratings.”

In this leaderboard ranking, Gen-4.5 has an Elo score of 1,247 points, marginally ahead of Google Veo 3 and Kling 2.5 from Kuaishou. Then came the latest versions of the Sora model developed by OpenAI. Here too, the lead is small enough for Runway to claim that, in this specific ranking, it is in the lead position.

Topping that list gives Runway an answer when customers ask how its results compare to better-known alternatives.

Where the model still breaks

In spite of the improvement in fidelity, Gen-4.5 is still fraught with classic AI video failures.

Runway itself points out problems in causal reasoning and object permanence—the feeling that events are in logical order and that objects are in the right place at the right time. In some scenarios, doors open without anyone reaching for them. 

In others, objects are removed or reinserted between scenes, or complicated stunts are pulled off in impossible ways. Viewers who have carefully inspected early versions for glitches point out that Gen-4.5 is better than Gen-3 and Sora versions at first, but it still stutters on temporal and physical logic when the tasks get complicated.

In practice, this matters for those attempting to employ AI video beyond attention-grabbing clips. Brand teams and game studios are interested in such areas as shot continuity, characters’ on-model behavior, or minute physical details that hold a scene together. Even if the model occasionally breaks and forgets certain details, it can still be put to use in pre-production stages.

Hardware and Enterprise Stakes

Underneath this creative layer, Gen-4.5 is also very much a hardware story. Runway says it built and launched this model solely on Nvidia GPUs, optimized for the use of the latest Hopper and Blackwell chips. Nvidia CEO Jensen Huang has described this effort as an application of their latest accelerators for “world models”, systems that go beyond understanding what the world looks like, to understanding what it moves like.

From an enterprise perspective, this alignment makes sense. If Gen-4.5 can run well on the same GPU generations that companies are already renting from large cloud providers, it makes it easier to include video creation as another line item on current infrastructure budgets.

Whether for product video in retail, motion graphics in advertising, or cinematics in games. Preliminary pilots report Runway is a game-changer.

What Gen-4.5 really tests is whether a specialist video lab can set the pace in a field dominated by general-purpose AI giants, and whether small points on benchmark leaderboards correlate to real-world differences in saved time, avoided revisions, and quality of output. If it does, the balance has shifted a bit away from the biggest labs and toward a smaller company that’s been working toward this moment since Gen-1.