Runway Gen-3 Alpha

What it is

Gen-3 Alpha is a video generation model from Runway. Runway announced it on 17 June 2024 as the first of a new series of models, trained on new infrastructure the company built for large-scale multimodal training^[1]. It is not a standalone app so much as the model underneath several Runway tools: Text to Video, Image to Video and Text to Image, along with features such as Motion Brush, Advanced Camera Controls and Director Mode^[1].

How it was trained and what it controls

Runway trained the model jointly on videos and images rather than on video alone, and says the result improves on its predecessor Gen-2 in fidelity, consistency and motion^[1]. Training on both kinds of data means the model sees still images, which carry fine detail, alongside video, which carries movement, instead of learning motion from video footage by itself.

On the time axis, Gen-3 Alpha offers fine-grained temporal control, including transitions and keyframing^[1]. That control rests on highly descriptive, temporally dense captions: captions that describe not only what is in a shot but how it changes over time, so the model can be directed through a continuous shot rather than asked for a single moment^[1]. In terms of content, it generates expressive, photorealistic human characters and interprets a wide range of styles and cinematic terms^[1].

Release, pricing and safeguards

After a short closed alpha with select testers, Gen-3 Alpha became available to everyone on 1 July 2024^[5], and The Decoder reported on 1 July 2024 that it had become available to users^[2]. Using it required a paid subscription starting at $15 per month, with generation billed through a credit system: on the standard plan, 625 credits equated to 62 seconds of generation^[5]. Extra credits could be bought in minimum blocks of 1,000 for $10^[5].

The model shipped with an in-house visual moderation system and C2PA provenance standards, and Runway also offered custom Gen-3 versions for media companies^[1]. Moderation and provenance marking are the two pieces that decide what the model will produce and how a generated file can be traced afterwards; offering custom versions puts the same model into the hands of organisations with their own production pipelines.

What was added later

In September 2024 Runway opened an invitation-only API for Gen-3 Alpha Turbo and added video-to-video to Gen-3 Alpha^[4]. On 22 October 2024, Act-One was released inside Gen-3 Alpha: it animates a generated character from a video of an actor's performance and voice, without motion-capture gear^[3]. For a video model, that extends control from describing a shot in words to using a real performance as a template.

Where it sits

Gen-3 Alpha follows Gen-2^[1], and its "series" framing mattered in practice: Runway shipped the first model of the series and then attached further capabilities — an invitation-only API, video-to-video, Act-One — to it over the following months^[4][3]. The name therefore covers a model that kept changing rather than a single frozen release.

As a product it also bundled delivery decisions with the model itself: credit-based paid access with a subscription floor^[5], built-in moderation, C2PA provenance marking, and custom versions for media companies^[1]. That combination — a general model, a metered way to use it, and visible content safeguards — is what made Gen-3 Alpha something a studio or a broadcaster could adopt, and not only a demo for individual users.