Runway (Gen-3, Gen-4, Aleph, Act-Two)

Runway is the oldest video-focused company in this document. It was founded in New York in 2018) by Cristóbal Valenzuela, Anastasis Germanidis, and Alejandro Matamala Ortiz, the three of them met at NYU's Interactive Telecommunications Program (ITP), the storied graduate program that has produced an unusual number of creative technologists. Runway predates the diffusion modelDiffusion modelA generative model that learns to produce images by training a neural network to reverse a noise-adding process. Generates by starting from random noise and iteratively denoising to produce a clean image. era entirely; the company started by building tools that helped creative professionals use existing AI techniques (style transfer, inpaintingInpaintingEditing a specific masked region of an image while preserving the rest. The classic AI editing approach., segmentation) and gradually evolved into a video generation lab as the underlying models got good enough to make that worthwhile.

Patrick Esser, who we have met repeatedly in this document as one of the original Stable Diffusion authors, was a researcher at Runway during the period when the latent diffusionLatent diffusionDiffusion that operates in a compressed 'latent' space produced by a VAE, rather than directly in pixel space. Dramatically reduces compute cost. Used by essentially every modern image and video model. paper was being written. The Munich-CompVis group and Runway were collaborators on that paper before they all went their separate ways, Rombach to Stability and then Black Forest Labs, Esser to Runway and then Black Forest Labs, the rest of the cast scattered. The fact that Runway was part of that collaboration matters because it gives them deep technical credibility, Runway is not just a product company, they are a serious research lab with researchers who have shipped foundational papers.

Runway has raised over $300 million across multiple rounds from Google, NVIDIA, Salesforce, and Felicis. The company was valued at around $1.5 billion in early 2024 and is presumably higher now. Chief Executive Officer (CEO) Cristóbal Valenzuela has positioned Runway as the AI-native answer to Adobe, a full creative suite where AI is built into every step rather than bolted onto existing tools. Lionsgate, the Hollywood studio, has partnered with Runway to develop custom models trained on their licensed content. Getty Images has done similar. This positions Runway uniquely as the AI video lab with the strongest entertainment industry relationships.

Gen-1 and Gen-2 (early 2023)

The original Runway video models. Gen-1 was video-to-video, you uploaded a video and it transformed it stylistically. Gen-2 was the first proper text-to-videoText-to-videoGenerating video from text alone, with no starting image. model from Runway. Both were early diffusion models that bolted temporal layers onto image-model backbones. Quality by today's standards was rough, short clips, limited motion, frequent artifacts, but at the time they were the best generally available video generation tools.

Gen-3 Alpha (June 2024)

A major upgrade. Gen-3 AlphaAlpha (LoRA alpha)A scaling factor that controls how strongly a LoRA's learned change is applied, effectively its volume knob, tuned separately from rank. was Runway's first model trained from scratch on video data with spatiotemporal attentionSpatiotemporal attentionAttention that operates across both spatial dimensions (within frames) and temporal dimensions (across frames). Essential for high-quality video generation. (rather than bolting time onto an image model). Quality jumped significantly. This was the version that put Runway competitively close to Sora during the period when Sora was still in research preview.

Gen-4 (early 2025)

Runway's current flagship video model. Improved consistency, better motion, longer clips, native 1080p output. Gen-4 Turbo is a faster, cheaper variant for iteration. Gen-4.5 is a higher-quality variant introduced in late 2025.

Aleph (July 31, 2025)

This is the model that distinguishes Runway from its pure-generation competitors. Aleph (the name comes from the Hebrew letter, suggesting 'beginning' or 'foundational') is a video-to-video editing model that transforms existing footage based on text prompts. You upload a video, you tell Aleph what to change ('remove the person on the left,' 'change the season to winter,' 'add a sunset in the background,' 'make this a low-angle shot'), and Aleph regenerates the video with your changes applied while preserving everything you did not ask to change.

Aleph is not the only video editing model in the field, but it is the one most closely integrated into a professional creative workflow. The architecture is based on Gen-4 with additional training for in-context editing, the input video is passed to the model as conditioningConditioningAny extra input that steers generation beyond the prompt, such as a depth map, pose, edge map, or reference image., and the model has been trained on enormous numbers of (original video, instruction, edited video) triplets. The use cases Runway emphasizes include adding VFXVFX (Visual Effects)Imagery created or altered beyond what was captured on camera, now increasingly generated., changing seasons or weather, relightingRelightingChanging the lighting on a subject after it was captured or generated, one of the harder problems in compositing. scenes, replacing or removing objects, generating new camera angles from the same source footage, and predicting the next shot in a sequence. For post-production work, Aleph is one of the most powerful tools currently accessible to anybody.

Act-Two and the performance capture line

Act-Two is another Runway model worth mentioning briefly. It is a 'performance capture' model that takes a driving video of a person acting and applies that performance to a different character. You record yourself doing a scene, and Act-Two regenerates the scene with a different character (an animated figure, a different actor, a creature) performing the same actions and saying the same lines. It is an AI-native version of motion capture, but without the physical capture rig.

Runway sits in an interesting place in the field. They have less raw compute than Google or OpenAI, but they have more focused expertise on the specific problems of professional video work, editing, performance capture, multi-shot consistency, the whole post-production pipeline. Their bet is that the future of video generation is not just 'type a promptPromptThe text description you provide to a model to specify what you want it to generate. and get a clipCLIPA text encoder developed by OpenAI in 2021 that learns to align text and images in a shared embedding space. Foundation of most text-to-image models from 2022 onward.' but a full creative environment that integrates generation with editing, performance, and refinement. As of 2026, this bet appears to be paying off, Runway is one of the few video model companies generating substantial revenue from professional creative customers, and the partnerships with Lionsgate and Getty are exactly the kind of relationships that defensibly differentiate them from the pure-model competitors.

On May 21, 2026, Runway released Aleph 2.0 alongside a new product called Edit Studio. Where the first Aleph proved that in-context video editing was possible, Aleph 2.0 made it production-shaped. It handles clips up to 30 seconds at 1080p and, more importantly, propagates a single reference-frame edit across multiple shots. You fix one frame, change the wardrobe, remove a sign, relight the scene, and the change carries through the sequence rather than forcing you to repeat the instruction shot by shot. That consistency across a cut is the exact problem that separates a demo from an edit bay.

Then on July 23, 2026, Runway shipped a Media Router on its developer platform. It automatically selects the best image, video, or audio model for a given request based on whether you prioritize quality, speed, or cost, and it is explicitly not limited to Runway's own models. This is a notable admission from a company that built its name as a model maker: it is a bet that in a crowded field, the durable position is to be the layer that routes work to whichever model is best right now, not to insist your own model always wins. It puts Runway in the same conceptual territory as fal and the other aggregators covered later in this book.

The commercial side moved just as fast. On June 11, 2026, Lionsgate took an equity stake in Runway and expanded a partnership that began in 2024, with plans to draw on existing franchises, John Wick among them, for AI-assisted short-form episodic series. On July 1, 2026, Runway announced a creative partnership with Bertelsmann spanning RTL and Fremantle, BMG, and the group's marketing businesses. Read together, these deals show Runway trying to convert model access into standing enterprise relationships, the kind of contracts that survive the next model release from someone else.

Runway was founded in 2018 by three people who met at NYU Tisch's Interactive Telecommunications Program: Cristobal Valenzuela, the Chilean CEO, Alejandro Matamala, the Chilean chief design officer, and Anastasis Germanidis, the Greek CTO. The founding line was to build new tools for human imagination, which is a real philosophical stance, not a tagline: Runway framed generative models as instruments for artists rather than as a step toward artificial general intelligence, and it planted itself in Manhattan on purpose, closer to the film, art, and design worlds it wanted to serve than to Silicon Valley.

The research pedigree is unusually deep for a company this size. In August 2022 Runway co-released Stable Diffusion, the open-source text-to-image model that seeded much of this entire field, working with the CompVis group at LMU Munich and with Stability AI's compute. It is a genuine irony worth sitting with: Runway helped open the floodgates on open image generation and now competes in the closed, premium video market those floodgates created. The money followed the research. Runway moved from a two million dollar seed through Series A, B, and C rounds to a 2023 raise that valued it around 1.5 billion, and then a 2025 round led by General Atlantic that put the valuation above three billion. Culturally it runs the AI Film Festival, whose submissions grew from a few hundred in 2023 to more than six thousand in 2025, and it partners with the Tribeca Festival, which is exactly the kind of legitimacy a base-model leaderboard cannot buy.

Runway's real edge is not raw generation, it is control over footage that already exists. Aleph, its in-context editing model, will relight a scene, add or remove an object, or generate a new camera angle of something you already filmed, all from a natural-language instruction, and that is where professional post-production money actually sits. Around it is a mature, integrated toolset rather than a single generate button: performance capture through Act-One and Act-Two, consistent characters and objects across shots through Gen-4, the Frames image model feeding the pipeline, and a genuine claim to Hollywood adoption, with Runway tech used on Everything Everywhere All at Once and to edit The Late Show. The workflow itself is the moatMoatA durable competitive advantage that protects a business from being copied or undercut, such as workflow depth, proprietary data, or brand and community..

The weakness is the mirror image. On any given week, the single highest-fidelity raw clipCLIPA text encoder developed by OpenAI in 2021 that learns to align text and images in a shared embedding space. Foundation of most text-to-image models from 2022 onward. is more likely to come from Google's Veo, Kuaishou's Kling, or ByteDance's Seedance than from Runway's own base model, and Runway is structurally out-compute-d by hyperscalers and mega-funded rivals if the game stays about scaling foundation models. Its answer, quietly selling access to competitors' models inside its own product, is a smart hedge and an admission at the same time: it concedes it may not own the best engine and makes itself partly dependent on the rivals it routes to. Consumer mindshare for just make me a video may also drift to Sora and Veo, which sit on far larger funnels, leaving Runway the narrower but higher-value professional lane.

The plans ladder cleanly. A free tier hands you a one-time credit grant, enough to evaluate and no more. Standard, around twelve dollars a month billed annually, is the real entry point: a monthly credit budget, every image and video model including Gen-4.5 and Aleph plus the third-party Veo and Nano Banana options, 4K upscalingUpscalingIncreasing an image or video's resolution after generation, often with a model that invents plausible fine detail rather than simply stretching pixels., and no watermarks. Pro roughly doubles the price and adds custom voice for lip sync and text to speech with a much larger credit pool, Max steps up again for heavy volume with credit rollover and first access to new models, and Enterprise adds single sign-on, analytics, teamspaces, and the APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface. for programmatic generation.

The way to actually get value out of it is to stop treating Runway as a slot machine for clips and start treating it as an editing and orchestration hub. Lock a character or a look first, generate candidate shots across whichever models win that particular shot, then refine, relight, re-angle, and composite with Aleph and Act-Two rather than betting everything on one-shot text to video. In short, it is strongest when you already have footage or a locked reference and you want control, and weakest when you just want the single prettiest clipCLIPA text encoder developed by OpenAI in 2021 that learns to align text and images in a shared embedding space. Foundation of most text-to-image models from 2022 onward. with no editing, which is the moment to route out to a frontier base model instead. The forward bet reads the same way: Runway is trying to become the professional creative operating system for generative media, adding an orchestration layer over every model and pushing into interactive world models, so that owning the creator's workflow outlasts owning any single engine.