Runway Gen-3 Alpha
Last reviewed
May 17, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v2 · 4,966 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 17, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v2 · 4,966 words
Add missing citations, update stale details, or suggest a clearer explanation.
Runway Gen-3 Alpha is a generative AI video model developed by Runway (the trade name of the New York based company Runway AI, Inc.). It was unveiled on June 17, 2024 as the first member of a new family of foundation models trained on what Runway described as a freshly built infrastructure for large scale multimodal training, and rolled out to paying subscribers on the Runway web app shortly after the announcement.[^1][^2][^3] Gen-3 Alpha generated 720p clips of roughly five or ten seconds in length from text prompts, still images or short video references, and was widely seen as a major step up from the company's earlier Gen-2 system in fidelity, motion realism and prompt adherence.[^2][^3] Within a few weeks Runway added an image to video mode on July 29, 2024[^4] and an Extend feature for stitching together longer sequences, and a faster and cheaper sibling model called Gen-3 Alpha Turbo followed in mid 2024.[^5][^6]
Gen-3 Alpha arrived during a stretch of intense competition in text to video generation. OpenAI's Sora had been previewed in February 2024 but was still gated to a tiny group of artists and red teamers, while Chinese systems such as Kling and consumer focused tools such as Luma's Dream Machine and Pika were beginning to ship to the public.[^2][^7] Runway's pitch with Gen-3 Alpha was that filmmakers and serious creative users could actually access a state of the art video model on a credit based subscription, and the model quickly became a fixture in commercial work, music videos, viral short films and stage visuals.[^7][^8] The launch was also followed in late July 2024 by a 404 Media report on a leaked internal training data spreadsheet, allegedly listing thousands of YouTube channels and films used as references for the model, which sparked a long running discussion about training data sources for video models.[^9][^10][^11] Gen-3 Alpha remained Runway's flagship video model until it was eclipsed by Runway Gen-4 on March 31, 2025.[^12][^13]
| Field | Value |
|---|---|
| Developer | Runway (Runway AI, Inc.) |
| Type | Text to video, image to video and video to video diffusion model |
| Announced | June 17, 2024[^1] |
| Public access | July 2024 (paid subscribers, Runway web app)[^2][^3] |
| Image to Video release | July 29, 2024[^4] |
| Architecture | Joint text and video diffusion transformer (details limited; see below) |
| Parameters | Not publicly disclosed |
| Training compute | Runway's internal research cluster (size not disclosed)[^1] |
| Output resolution | 720p (native)[^2] |
| Base clip length | 5 seconds and 10 seconds[^2] |
| Variants | Gen-3 Alpha, Gen-3 Alpha Turbo[^5][^6] |
| Successor | Runway Gen-4 (March 31, 2025)[^12] |
| License | Proprietary, web platform and API |
Runway was founded in 2018 by Cristobal Valenzuela, Alejandro Matamala and Anastasis Germanidis, three graduates of New York University's Tisch School of the Arts who met at the Interactive Telecommunications Program.[^14] The company started as a browser based suite of small machine learning tools for creators and gradually built a reputation for shipping research grade models inside a usable interface. Runway was also a co-author, with researchers at LMU Munich, of the influential 2022 paper on latent diffusion models that became the basis for Stable Diffusion.[^14]
The company's first generation of branded video models, called the Gen series, traced an arc from controlled video editing to open ended generation:
In parallel, the AI video field was moving fast. OpenAI had teased Sora in February 2024 with sample clips that were noticeably longer and more cinematic than anything publicly available, putting pressure on every other lab to demonstrate their next generation model.[^15] Runway's response was Gen-3 Alpha, announced about four months after the Sora teaser and shipped to paying users while Sora itself was still locked behind a closed program.[^2][^15]
Runway unveiled Gen-3 Alpha on June 17, 2024 with a research blog post titled "Introducing Gen-3 Alpha: A New Frontier for Video Generation" and a reel of sample clips, calling the model "the first of an upcoming series of models trained by Runway on a new infrastructure built for large-scale multimodal training."[^1] The announcement emphasised improvements over Gen-2 in three areas: fidelity (textures, lighting, photoreal humans), consistency (motion across the clip) and motion (responsive, physically plausible movement).[^1][^2]
Gen-3 Alpha was the first time a high quality general purpose video model was actually available to anyone with a credit card, on a normal web app, with reasonable rendering times.[^2][^3] Sora was widely covered but not usable; Pika, Luma Dream Machine and Kling all offered free or cheap access but with weaker motion and lower fidelity at launch.[^7] For roughly the second half of 2024, Gen-3 Alpha sat in an unusual position as both a frontier research demo and a commercial product, and much of the public's first impression of what serious AI video looked like came from Runway clips circulating on social media.[^7][^8]
Coverage at launch was overwhelmingly positive. The Verge described the system as a noticeable jump in quality and a sign that Runway had "closed a lot of the gap with Sora" without restricting access to a hand picked group of testers.[^2] TechCrunch focused on the new controls Runway exposed in Gen-3 Alpha, including more granular camera direction and temporal control through "highly descriptive, temporally dense captions."[^3] PetaPixel highlighted Gen-3 Alpha's ability to produce realistic human characters with a range of emotions and actions, a capability that Gen-2 had largely lacked.[^16]
Runway initially rolled the model out to a small group through a waitlist and the company's creative partners program in mid June 2024, then opened it to paying subscribers on the Runway web app over the following weeks.[^1][^2] At launch the model supported only text to video; image to video and other modalities followed within roughly six weeks.
On July 29, 2024 Runway opened Gen-3 Alpha Image to Video to all paying users, allowing a single still image to be used as the first frame of a clip with an optional text prompt for additional guidance.[^4] This update was widely seen as the key bridge between separate AI image tools (such as Midjourney or Stable Diffusion) and Runway's video model: users could compose a key frame in a high control image tool, then hand it to Gen-3 Alpha to animate.[^4]
Within weeks Runway followed with Gen-3 Alpha Turbo, advertised as roughly seven times faster than the original Gen-3 Alpha for image to video generation and approximately half the credit cost.[^5][^6] VentureBeat reported that the Turbo variant could generate AI videos "faster than you can type" and was rolled out across all Runway plans including free trials.[^5] Turbo is image to video by design (it requires a source image) and trades a small amount of fidelity in fine motion for much faster iteration.[^5][^6]
Both models share the Gen-3 Alpha backbone but tune the inference path differently. Runway has not published technical details of the Turbo distillation procedure.
| Variant | Released | Key traits | Typical use |
|---|---|---|---|
| Gen-3 Alpha | June and July 2024 | Highest quality output in the family; slowest and most expensive per second; supports text to video, image to video and video to video | Hero shots, music videos, key sequences in commercials, festival films |
| Gen-3 Alpha Turbo | 2024 (after Image to Video update) | Roughly 7x faster and 50% cheaper per credit; image to video required; small loss of fidelity in fine motion compared to Alpha | Iteration, previs, social posts, large batch generation, Act-One performance transfer |
Runway has been deliberately sparse about Gen-3 Alpha's internals, in line with its general policy across the Gen series. The publicly stated facts come from the company's research blog post and from interviews with Cristobal Valenzuela and Anastasis Germanidis around the launch.[^1][^17]
Runway pitched Gen-3 Alpha as the first model on its "new infrastructure built for large-scale multimodal training," framing it as the start of a series of "general world models."[^1] At launch the practical difference was simply that Gen-3 Alpha was a much more coherent video model than Gen-2 in basically every dimension creators cared about.[^2][^7] This article does not speculate about internals (parameter count, attention layout, tokeniser choices, denoiser schedule, etc.) that Runway has not publicly disclosed.
Gen-3 Alpha generated 720p video at variable frame rates, with 24 frames per second as a common output for clips that mimicked traditional cinema.[^2] Press reviews described three headline improvements over Gen-2:[^2][^3][^16]
The model still struggled with hands and very fast multi object interaction, and produced occasional hallucinations such as duplicated limbs or unstable text on signs, but the failure rate per generation dropped substantially compared with Gen-2.[^2][^3]
The launch mode for Gen-3 Alpha was straightforward text to video: a user wrote a prompt and the system generated a clip from scratch at 5 or 10 seconds.[^2] Runway emphasised that Gen-3 Alpha had been trained with highly descriptive, temporally dense captions, which let the model interpret prompts that referenced changes over time (such as "a slow zoom in" or "the candle gutters and goes out") with more precision than Gen-2.[^1][^3]
On July 29, 2024 Runway released Gen-3 Alpha Image to Video, which let users supply a still image as the first frame, with or without a text prompt that described how the scene should evolve.[^4] This made Gen-3 Alpha much more controllable in production workflows: users typically generated a key frame in a separate image model such as Midjourney or Stable Diffusion, then handed that frame to Gen-3 Alpha to animate.[^4][^7] In the Turbo variant, image to video became the primary supported mode.[^5][^6]
Gen-3 Alpha also accepted short reference video clips and could re-style or continue them, building on the lineage of Gen-1's video to video work.[^1][^7]
Gen-3 Alpha exposed explicit camera controls in addition to the text prompt: the user could select a camera move (zoom in, zoom out, pan, tilt, dolly, orbit), set the intensity of that move and combine multiple moves in a single shot, in a feature Runway described as "Director Mode."[^1][^3] This turned Gen-3 Alpha into a usable tool for previs and storyboarding, since users could iterate on a shot composition without rewriting the prompt every time.
Gen-3 Alpha's base output length was 5 or 10 seconds.[^2] To produce longer sequences, users could apply the Extend feature, which conditioned a new generation on the last frame of the previous clip and continued the action. In practical workflows, two extensions on a 10 second base clip produced sequences in roughly the 16 to 18 second range, although quality and continuity tended to drift in later extensions.[^6][^7] For longer sequences most filmmakers cut between many short Gen-3 Alpha clips rather than relying on a single extended generation.
Gen-3 Alpha preserved style across a clip more reliably than Gen-2, but did not yet support consistent character transfer across cuts — a feature that arrived only with Gen-4 and its References system in March 2025.[^12][^13]
In October 2024 Runway released Act-One on top of Gen-3 Alpha. Act-One let users record a short video of a real performer, then drive an AI generated character with the same facial expressions, head movement and gesture.[^18] It was widely seen as the most credible character animation feature in any consumer AI video tool at that point.[^18] Act-One later extended to Gen-3 Alpha Turbo for cheaper iteration.
Runway also added a Lip Sync feature on top of Gen-3 Alpha, which took an existing clip of a character and re-rendered the mouth area to match an uploaded audio track or text to speech output. This was not real time and did not produce native synchronised audio inside the model itself; the audio still had to be generated or recorded separately and then aligned. Native audio generation came to Runway only with later models.
Gen-3 Alpha launched into a crowded field. The table below compares it with the major contemporary text to video systems as of mid 2024 to early 2025.
| Model | Developer | First public release | Max clip length | Native audio | Notes |
|---|---|---|---|---|---|
| Gen-3 Alpha | Runway | June 2024 (announce), July 2024 (paid)[^1][^2] | ~10s base, longer via Extend[^2][^6] | No (Lip Sync later) | Realistic motion, public access from launch |
| Sora (original) | OpenAI | Announced February 2024[^15] | Up to 60s in demos | No (original) | Closed alpha through most of 2024 |
| Veo | Google DeepMind | May 2024 (Veo 1 demo)[^19] | Demos at 60s | No (Veo 1) | Limited Vertex AI preview |
| Veo 2 | Google DeepMind | December 2024 | ~8s typical | No | High fidelity, restricted access |
| Veo 3 | Google DeepMind | May 2025 | ~8s typical | Yes | Native audio, public via Gemini |
| Kling (1.0) | Kuaishou | June 2024 | Up to ~2 minutes | No (initial) | Long clips, initially geo gated |
| Luma Dream Machine | Luma AI | June 2024 | ~5s base, extendable | No (initial) | Free tier, lower fidelity at launch |
| Pika | Pika Labs | Late 2023 | ~3 to 10s | No (initial) | Fast iteration, lower fidelity than Runway |
| Hailuo (Video-01) | MiniMax | September 2024 | ~6s | No (initial) | Strong realism in human motion |
For roughly the second half of 2024, Gen-3 Alpha was the most plausible answer to the question "how do I actually get state of the art AI video right now," and that role only began to slip in late 2024 as Sora Turbo, Veo 2 and Kling 1.5 reached more users.[^7][^8]
Reception of Gen-3 Alpha was unusually positive for a Runway release. The Verge described the model as a noticeable jump in quality over Gen-2 and observed that Runway had "closed a lot of the gap with Sora" without limiting access to a hand picked group of testers.[^2] TechCrunch emphasised the improved controls and described the model as offering "high-quality, photorealistic clips."[^3] PetaPixel focused on the photorealistic human characters and the range of actions and emotions the model could express, contrasting that with the limited human rendering of Gen-2.[^16] VentureBeat treated the launch and the subsequent Turbo release as a meaningful inflection point in commercial AI video generation, with Turbo making the technology accessible to free tier users for the first time.[^5][^20] Tom's Guide characterised Gen-3 as "a big leap forward" for accessible AI video.[^21]
The community on Twitter/X and on Runway's own Discord pushed thousands of clips through the model in the first weeks, with a lot of attention going to filmmakers who had already been doing notable work with Gen-2 and Pika and now had a substantially better tool.[^7][^8] Clips set in physically grounded contexts (cars, weather, food, architecture) looked unusually clean for AI video; clips with crowds, fast multi object interaction and complex hands were where Gen-3 Alpha's seams still showed. A meaningful share of films submitted to Runway's AI Film Festival in 2024 and 2025 were produced primarily on Gen-3 Alpha and Gen-3 Alpha Turbo, and the model became a fixture in commercial work, music videos, viral short films and stage visuals during this period.[^7][^8]
In late July 2024, the technology publication 404 Media published a report based on what it described as a leaked Google Sheet from Runway's internal "Jupiter" project — the codename for what was released publicly as Gen-3 Alpha.[^9][^10] According to the report, the spreadsheet contained thousands of rows listing YouTube channels, films, TV shows and individual videos that had allegedly been used as references or training data for Gen-3 Alpha. The listed YouTube channels included major media and entertainment companies (The New Yorker, VICE News, Pixar, Disney, Netflix, Sony) and individual creators (Casey Neistat, Sam Kolder, Benjamin Hardman, Marques Brownlee).[^9][^10][^11] 404 Media reported that Runway had scraped videos using the open source tool yt-dlp with proxied IP addresses to avoid being blocked by YouTube, and that the spreadsheet also listed pirated film and television content.[^9][^11] 404 Media stated it had verified that the document originated inside Runway.[^9]
YouTube's terms of service prohibit downloading content without permission, and the practice as described would constitute a violation.[^9][^10] Runway did not publicly publish a line by line response. Cristobal Valenzuela addressed the story in passing in interviews and on social media, framing the company's data practices in terms of fair use, in-house pipelines and partnered datasets, but did not confirm or deny that the spreadsheet was a real internal document.[^9][^10] Runway has not released a model card itemising training sources for Gen-3 Alpha. The report fed into a wider 2024 and 2025 conversation about training data for video models, and other companies in the space, including OpenAI for Sora and several open source video efforts, faced similar questions about whether YouTube and movie content had been used. The controversy did not visibly slow Gen-3 Alpha's commercial uptake, but it has remained a frequently cited reference point in academic and policy discussions about training data transparency in generative AI.
In September 2024 Runway announced a first-of-its-kind partnership with film studio Lionsgate to create a custom AI video model trained on Lionsgate's catalog of more than 20,000 film and TV titles, including The Hunger Games franchise.[^22][^23] The deal was for a jointly managed, exclusively licensed custom model rather than for general use of Lionsgate content in Runway's public Gen series. Lionsgate executives framed the deal as a way to augment, not replace, existing creative workflows, with expected use cases including pre-visualization, storyboard generation, background and environment creation, and certain categories of visual effects work such as action sequences and explosions.[^22][^23] Subsequent reporting in 2025 suggested the partnership had run into difficulties, with one source telling The Wrap that "the Lionsgate catalog is too small to create a model."[^24]
Runway has used a credit based subscription model since before Gen-3 Alpha. Credits cap how much video a user can generate per month at each tier, with extra credits available as add-on top ups. The pricing below reflects the structure that was in place when Gen-3 Alpha was the company's flagship model in 2024.
| Plan | Monthly price (annual billing) | Monthly credits | Notes |
|---|---|---|---|
| Basic | Free | 125 one time | Limited generations; watermarked exports for some content |
| Standard | $12 | 625 | Unlocked exports; access to Gen-3 Alpha and Turbo |
| Pro | $28 | 2,250 | Higher resolution exports, longer clip features |
| Unlimited | $76 | Unlimited generations in Explore Mode plus 2,250 fast credits | Designed for heavy users; render queue prioritisation |
| Enterprise | Custom | Custom | Dedicated GPUs, white labeling, REST API SLA, custom training options |
Gen-3 Alpha was billed at roughly 10 credits per second of generated video for the Alpha model and roughly 5 credits per second for the Turbo variant, with Extend generations billed at the same per second rate as the base mode.[^25][^5] A Standard plan with 625 credits therefore worked out to roughly 62 seconds of Gen-3 Alpha video or about 125 seconds of Gen-3 Alpha Turbo video per month before having to buy top ups. Runway has revised credit pricing several times over the life of the model, particularly as Gen-4 became the flagship, and as of 2026 Gen-3 Alpha and Turbo are billed at lower credit costs than Gen-4 and later models.
Runway opened a developer API for Gen-3 Alpha Turbo (image to video) in September 2024, allowing third party tools and pipelines to call the model programmatically. The general Gen-3 Alpha model was made available through the same API surface for paying customers, with rate limits and pricing aligned to the subscription tiers.
On March 31, 2025 Runway announced Gen-4 as the successor to the Gen-3 Alpha family.[^12][^13] The headline feature was the References system, which lets a user upload one or more reference images that the model treats as visual anchors for character, object and scene identity across separate generations — without any fine-tuning or retraining of the model.[^12][^13] This addressed Gen-3 Alpha's most widely criticised weakness: in longer narrative work, the same character would visibly drift in appearance from one Gen-3 clip to the next.[^7][^12]
Gen-4 also offered:
Gen-4 Turbo followed on April 7, 2025 as a faster and cheaper variant of Gen-4. Runway said Gen-4 Turbo could generate a 10 second video in roughly 30 seconds, making it about five times faster than the standard Gen-4 model while maintaining similar quality. Gen-4 Turbo was priced at roughly 5 credits per second, the same per second rate as Gen-3 Alpha Turbo, and was rolled out across all Runway plans.
The Gen-4 release ended Gen-3 Alpha's role as the company's flagship video model. Gen-3 Alpha and Gen-3 Alpha Turbo remained available on the Runway platform for users who preferred the older model's specific look or who had built workflows around its API, but new feature work shifted to Gen-4 and later models.[^12][^13]
On July 25, 2025 Runway introduced Aleph (initially referred to in some Runway documentation as "Gen-4 Aleph"), described as a state-of-the-art in-context video model rather than a traditional generative video model.[^26] Unlike Gen-3 Alpha and Gen-4, which generate video from a prompt or a reference image, Aleph operates on existing video footage and applies natural language edits to it.[^26] Supported operations include adding or removing objects from a scene, transforming the appearance of objects already present, changing lighting and time of day, altering the visual style of the footage, and generating alternate camera angles from a single shot.[^26]
Aleph was made available to all paid Runway users on launch and was added to the Runway API shortly afterwards, becoming an important post-production tool for editors and visual effects workflows.[^26] Although Aleph is named for the Hebrew letter Aleph, it is not part of the Gen-3 Alpha family; the relationship is best understood as a parallel branch of Runway's product line that focuses on editing rather than generation.
As of mid 2026, Gen-3 Alpha and Gen-3 Alpha Turbo remain available on the Runway platform but are no longer the company's flagship video models. New feature work has been focused on Gen-4 and its successors, with Aleph occupying the editing branch of the product line.
Several aspects of Gen-3 Alpha have had a durable impact on the AI video field. Accessibility set a new bar: Gen-3 Alpha demonstrated that a frontier quality video model could be shipped as a paid web product within months of being announced, rather than spending a year or more in closed alpha, which contributed to the shift in industry expectations that pushed Sora, Veo and Kling toward broader public release in late 2024 and 2025. The combination of explicit camera control and Act-One performance transfer made Gen-3 Alpha the first AI video model that some filmmakers described as a usable production tool rather than a curiosity.[^18] The 404 Media Jupiter report became a touchstone in subsequent debates about training data disclosure, model cards and consent for AI video models, and is regularly cited in academic and policy work on the topic.[^9][^10][^11] Gen-3 Alpha was also the first AI video model to see meaningful, repeated use in commercial work — music videos, advertising, stage visuals, late night TV segments and film pre-visualisation — at a scale beyond isolated demos.[^7][^8][^22]
In retrospect Gen-3 Alpha is generally framed as the model that took AI video generation from "interesting research demo" to "tool that working creators will actually pay for," even though its specific quality ceiling has been surpassed by Gen-4, Sora 2, Veo 3 and Kling 2.0 in the intervening years.