AI image and video generation: the model layer and the tools on top

Every AI video you have seen this year came from a company that trained the model, or one that built a product on someone else's. The line is blurring: model makers ship their own apps, and app makers let you pick other companies' models.

We checked 35 companies against their own websites and pricing pages on 25 September 2026. This is the map, grouped by what each company sells.

The short version

• →The market splits into four groups: model makers selling closed models, labs releasing open weights, creative platforms and editors built on top, and avatar and presenter tools.
• →The tools on top are becoming model-agnostic. At least 10 of the 35 companies we checked offer other companies' models inside their own product, from Adobe and Krea to HeyGen and Synthesia.
• →Big names can leave. OpenAI shut the Sora app in April 2026 and scheduled the Sora API to close on 24 September 2026.

Model makers selling closed models

Runway says it is "building foundational Real-World Intelligence", and sells an app for creative teams and an API for developers. In June 2026 Lionsgate took an equity interest in Runway. Luma AI, based in Palo Alto, sells its Ray video models through an app for creative teams and an API. Kling sells video and image generation to creators through its app and to developers and enterprises through its API. MiniMax sells its video models through the Hailuo app and an API, alongside language, speech and music models. Google Veo is Google's video model, "designed to empower filmmakers and storytellers", available through the Gemini app, Flow and the Gemini API. Midjourney sells image and video generation by subscription through its web app and Discord.

Seedance, from ByteDance, is an audio and video generation model sold to individuals and enterprises and used in ByteDance's own apps.

Moonvalley sells Marey, a video model "trained on licensed data", to filmmakers. In June 2026 Reka announced that it had joined forces with Moonvalley. The same release describes this both as "adding a team of AI researchers and engineers" and as "the merger", and Moonvalley's own site does not mention it, so we cannot say which reading is right.

OpenAI discontinued the Sora web and app experiences on 26 April 2026, and scheduled the Sora API to be discontinued on 24 September 2026. That date has passed, but OpenAI's help page still described the API shutdown in the future tense when we rechecked it on 12 October 2026, so we cannot confirm it has happened.

Labs releasing open weights

These labs publish model weights you can run yourself, and usually sell an API and commercial licences alongside.

Black Forest Labs, with labs in Freiburg and San Francisco, calls itself "the frontier AI research lab for visual intelligence" and sells its FLUX models by API and by licence. Stability AI sells image, video, 3D and audio models to enterprises and creative industries, with a free community licence for researchers, developers, small businesses and creators below a revenue threshold. LTX, which began as a division of Lightricks, builds open video and audio models and sells an API and a studio app. Genmo, which describes its work as "open world models", publishes its Mochi 1 video model under an open licence. Ideogram released open weights for Ideogram 4.0 and sells an app and API for teams and developers. MiniMax calls H3 "a general-purpose, omni-modal generative system" and publishes its weights under a community licence that excludes some territories, with an application form for the US, EU, UK and South Korea.

Creative platforms and editors built on top

Adobe Firefly generates and edits images, video, audio and designs using Adobe's own models and partner models. Krea calls itself "the world's most powerful creative AI suite" and offers its own models plus many third-party ones. Pika relaunched in September 2026 as a platform offering third-party models alongside its own. Recraft, headquartered in London, sells image and vector generation to "designers, creatives, sellers, and teams", and offers third-party models too. Leonardo.Ai, a creator platform with its own models and third-party ones, has been part of Canva since Canva announced its acquisition in July 2024.

Editing tools sit next to them. Descript edits video and audio by editing text. Captions sells an AI editor, and its company renamed itself Mirage in September 2025 while keeping the Captions product name. Kapwing sells collaborative editing for creative teams. VEED, headquartered in London, sells video creation "made for social" and lists third-party models alongside its own. OpusClip turns long videos into short clips. Invideo sells "the agentic video editor for serious creatives", built on third-party models. Adobe Premiere extends clips with Firefly.

Avatar and presenter tools

These companies sell finished video with a digital presenter, mostly to training, marketing and communications teams.

Synthesia, based in London, sells an "AI video platform for business" with third-party model integrations. HeyGen sells avatar video, translation and a video agent, with its own avatar model and third-party ones. D-ID sells videos and interactive avatars, and in September 2025 announced an agreement to acquire simpleshow, expecting to complete it in the fourth quarter of 2025. That expectation is now nearly a year old, and we found no confirmation that the deal closed. Colossyan, headquartered in Budapest, sells training video. Tavus, which calls itself "The Human Computing Company", sells real-time conversational video through an API. DeepBrain AI, headquartered in Seoul, sells AI Studios, which offers third-party video models. Elai.io sells video for learning and development and was acquired by Panopto in 2024. Wix acquired Hour One in 2025. hourone.ai no longer serves a secure page, and its plain web address redirects to a Wix video maker.

What to ask before you buy

1. 1.Whose model is this? If the product routes to other companies' models, ask which ones, and what happens to your work if one is withdrawn, as the Sora app was.
2. 2.What are you paying for? Convert seconds, credits and minutes into the cost of the output you need.
3. 3.What was the model trained on? Some vendors make licensed training data part of their pitch. Ask every vendor the same question.
4. 4.Can you run it yourself? Open weights change the answer for teams with their own infrastructure. Check the licence terms for your company's size and your country.

If you sell one of these products, expect these questions in the first reply. We look at the companies selling AI to media businesses in a companion piece, our AI in content and media landscape.

How we did this

We checked 35 companies against their own websites and pricing pages on 25 September 2026, and profile 31 of them here, described the way each describes itself; Adobe appears twice, for Firefly and Premiere. Before publication we rechecked every quoted self-description on 12 October 2026 and corrected the ones the companies had changed. Sora and Hour One appear only as events, and we left out two whose pages would not load for us. Ownership changes come from the companies' own announcements. Where a single source describes an event in two ways, we say so. The list covers the companies we examined, not the whole market.

If something here is out of date or wrong about your company, tell us and we will correct it with the source.