Runway Gen-4.5

What it is

Runway Gen-4.5 is a video generation model released by Runway on 1 December 2025. It turns text prompts into video clips, the task AI video tools have been built around. In its announcement Runway claimed state-of-the-art results in three areas: motion quality, prompt adherence and visual fidelity^[1]. Those three measure different things — how convincing the movement inside a shot looks, how closely the result follows what the prompt asked for, and how clean and detailed the individual frames are.

The model was not sold as a separate product. It was included in all paid Runway plans and available on the web and in the iOS and Android apps^[1].

Where it came from

Runway places Gen-4.5 at the end of a line that begins with Gen-1, which the company described as the first publicly available video generation model, released two years earlier^[4]. Between those two points the field moved from short clips with obvious artefacts toward output aimed at real production work.

According to Runway, the gains came from two places: pre-training data efficiency and post-training techniques^[4]. Pre-training is the long stage in which a model learns from large amounts of video; data efficiency refers to how much it can learn from a given quantity of that material. Post-training is the later stage in which a model is adjusted toward particular behaviour — following instructions, producing cleaner motion — typically using curated examples or human judgements about which outputs are better. Runway's claim was that both stages improved.

The model kept the speed and efficiency of Gen-4, the previous generation, and was offered at comparable pricing across all subscription plans^[4]. It was developed entirely on NVIDIA GPUs, and inference — the process of actually generating a clip when a user asks for one — runs on NVIDIA Hopper and Blackwell GPUs^[1].

How it performs

At launch, Gen-4.5 ranked first on the Artificial Analysis Text-to-Video leaderboard with 1,247 Elo, ahead of Google Veo 3 at 1,226, Kling 2.5 at 1,225 and OpenAI Sora 2 Pro at 1,206^[1]. Elo ratings, borrowed from chess, are built from many pairwise comparisons: raters see two outputs and pick the better one. A top position therefore means the model won more of those comparisons than its rivals; it is not an absolute measure of quality. The margins here were small, with the three nearest competitors sitting within a narrow band.

Runway also reported better physical accuracy: objects with realistic weight and momentum, and proper liquid dynamics^[1]. Physical plausibility is a specific weak point of video models, because a system that has learned from appearance alone can produce motion that looks right frame by frame but behaves wrongly over time — a ball that does not slow as it should, water that pours like smoke.

Limits

Runway listed the model's failures alongside its results: causal-reasoning errors, object-permanence problems, and what it called a "success bias", in which actions succeed too often^[1]. Causal reasoning means getting cause and effect right — understanding that one event produces another, rather than merely following it. Object permanence, a term borrowed from developmental psychology, refers to things continuing to exist and stay consistent when they leave the frame or are briefly hidden. The success bias is a data problem as much as a modelling one: if the video a model learns from mostly shows tasks being completed — doors opening, objects caught — the model will tend to make every attempt succeed, which reads as unrealistic.

Audio and editing

On 11 December 2025, Runway announced native audio generation for Gen-4.5, covering dialogue, sound effects and background audio, together with audio editing and multi-shot editing^[2]. By February 2026, TechCrunch described the model as having added native audio, long-form and multi-shot generation, character consistency and editing tools^[3]. Taken together, these move a video model from producing an isolated clip toward producing material that can be assembled into a sequence, with the same character and a continuous soundtrack.

Where it sits

Runway's first world model, GWM-1, is built on top of Gen-4.5^[2]. A world model, in the general sense, is a system that learns how an environment behaves and predicts what happens next in it, rather than only rendering what it is asked for. Video generation is one route toward that, because predicting the next frames of a scene requires some model of how the scene works. Building GWM-1 on Gen-4.5 made the video model the base layer for that line of work.