The distinction matters because every other major AI video model, including Sora, Runway, and Veo, currently relies on upscaling to reach 4K. That approach can introduce visual artefacts and adds a finishing step to the workflow. Kling’s native 4K generates every pixel at full resolution from the start, which the company says holds up better on close-up shots, fine textures, and detailed product work.

Kling cites two production credits to support the launch. The Chinese period drama Swords Into Plowshares used the model to build territorial maps and to compress a storm-sequence previs from two months to two weeks. House of David season one incorporated 72 AI-assisted shots, with Kling among the tools used. Both are previs and effects-support credits rather than primary photography, but they place Kling closer to professional production pipelines than most of its competitors.

Alongside the launch, Kling has opened a 4K Short Film Creative Contest with a $25,000 prize pool, 70,000 platform credits, and a screening event in South Korea for winning entries. Native 4K output is available now to existing 3.0 users.

For working creators, the practical takeaway is that the upscaling step many have built into their delivery workflow may no longer be necessary for Kling-generated content. For agencies and production companies handling premium ad work, the announcement removes one of the standing objections to AI video in 4K-mandated campaigns. Broadcasters will still need to assess codec, colour space, and HDR support before treating Kling output as mezzanine-ready, but on resolution alone the gap has closed.

The model, branded ChatGPT Images 2.0 on the consumer side, went live for ChatGPT and Codex users on 22 April, with the API following in early May. Microsoft has already rolled it out through its Foundry platform for enterprise use. On LM Arena’s image benchmarks it has taken the top spot, ahead of Google’s Nano Banana 2.

For anyone working in AI video, the interesting feature is not the upgraded text rendering, though that is genuinely useful for on-screen graphics and title cards. It is the multi-image batch generation in the Thinking mode variant. The model can produce up to eight panels from a single prompt while holding character identity, object placement and colour palette consistent across the set. That is, functionally, a storyboard tool.

It is also an image-to-video feed. Runway, Kling and most serious AI video pipelines now accept a sequence of reference images as the starting point for generation. Consistent keyframes have been the hard bit. A model that generates eight on-brand, on-character stills in one pass, with legible text inside them, removes a meaningful step from the workflow of anyone making episodic content, advertising, or vertical drama.

Pricing sits in the enterprise tier. Microsoft Foundry lists token-based pricing of $8 per million input tokens and $30 per million output tokens for gpt-image-2, with a cheaper mini variant at roughly a quarter of the output cost. Per-image cost lands somewhere between four and thirty-five cents depending on resolution and prompt complexity. The model supports output up to 4K on Foundry and adds stronger multilingual text rendering, which matters for anyone producing localised versions of the same asset.

The wider picture is harder to ignore. OpenAI announced the Sora shutdown on 24 March. The app closes on 26 April, the API on 24 September. A reported $1 billion Disney licensing deal collapsed with it. Bill Peebles, who led Sora, left OpenAI on 17 April. Four days later the company launched an image model with storyboard features. Internally the company has been shedding what it calls “side quests” to focus on enterprise revenue, and video generation, as it turns out, was one of them.

That leaves Google, Kling, Runway and a shrinking field of specialists as the serious options for AI video generation itself. OpenAI’s new position is one step upstream, at the pre-production layer, where the margins are better and the compute is cheaper.

For AI video producers, the practical takeaway is straightforward. If you are already building image-to-video workflows, gpt-image-2 is worth testing as a front end. It solves the two hardest problems in reference image generation, character consistency across a set and legible on-screen text, and it is priced as production tooling rather than consumer experiment. Whether it will hold that lead against Nano Banana and the next wave of open models is a separate question.

The absences are louder than the inclusions.

Runway ($860m raised, New York) returns as the default enterprise pick, consistent with recent mid-market spend data from YipitData that shows it as the only AI video generation platform meaningfully monetising at scale. Synthesia ($535m, London) holds its place as the enterprise avatar leader and is the only UK company on the list, which is worth a small flag for anyone tracking where European AI video is actually winning. HeyGen ($74m, Los Angeles) is the newcomer, and the number that jumps off the page is how little capital it has raised relative to its peers. On the list alongside Synthesia with roughly one-seventh the funding, HeyGen’s presence reflects the customer growth YipitData recorded over the past year, with mid-market customer counts growing 152 percent year on year to January 2026.

Luma is not on the list, despite a reported $900m Series C in November 2025 and a headline partnership with Wonder Project and AWS announced yesterday under the Innovative Dreams banner. Pika is not on the list. Higgsfield, Moonvalley, Genmo and OpenArt are all absent. Neither is Google’s Veo, Meta’s Movie Gen, or Microsoft’s video tooling. The Forbes list covers privately held companies only, which means the biggest incumbents in AI video by compute, distribution and user base are structurally excluded. The list tells you who’s winning among the challengers, not who’s winning overall.

Forbes’ framing for this year’s list explains the rest. Shrivastava writes that the judges rewarded companies that “productize AI in new and innovative ways and build sustainable businesses on top,” with the race shifting from raw model power to “ownership, infrastructure and influence.” The selection is weighted toward revenue, customer traction and commercial deployment rather than demo-reel capability.

Read through that lens, the video picks make sense. Runway has enterprise contracts. Synthesia has average contract values roughly three times HeyGen’s and a defensible enterprise base. HeyGen has fast-growing paid customer adoption. The companies that didn’t make it are, variously, still raising on valuation rather than revenue, still operating primarily as creator-facing tools, or in the case of the Chinese players, outside the Forbes geographic frame.

Three other names matter for AIVW readers. Black Forest Labs ($450m, Freiburg) is on the list for image and video, a German company that has quietly built a licensing business while Luma and Pika absorbed most of the attention. Midjourney makes the list having raised zero external capital and extended into video on the back of its own cash flow. ElevenLabs ($800m) is on for voice generation, a reminder that audio is where the commercial story is now tightening. Fal ($330m) provides inference infrastructure for a lot of the video tools that didn’t make the list themselves, which is its own quiet commentary.

The practical takeaway is straightforward. The gap between which AI video companies dominate press coverage and which ones show up on a list built around revenue and enterprise deployment is wider than most tool comparisons suggest. For creators and production professionals deciding where to place their time and paid subscriptions, the Forbes list is a useful corrective to marketing noise. The companies that get to keep shipping products are, in the end, the ones that figured out how to charge for them.

Source: Forbes AI 50, compiled by Forbes in partnership with Sequoia Capital and Meritech Capital, published 16 April 2026. Spend data: YipitData.

Readers with functioning memories may recall hearing something similar before. Asteria’s Marey model has been pitched as the first production-ready generative video tool trained on licensed data. Promise has positioned itself as the first AI-native studio working with Hollywood talent. Late Night Labs has talked up its hybrid production pipeline. Moonvalley, Wonder Dynamics, Cuebric, all have filed variations of the “first at scale” paperwork. The launch press release is becoming a genre in itself.

What’s actually on offer here is a virtual production stage at MBS Media Campus in Manhattan Beach, a partnership with Luma for what the release calls “production-grade AI tools,” and a first project, The Old Stories: Moses, a three-part companion to Wonder Project’s House of David. It stars Ben Kingsley and O-T Fagbenle, is written and directed by Erwin, and debuts on Wonder’s Prime Video subscription service this spring. In other words, an Amazon-backed studio making an Amazon-distributed show on AWS infrastructure, with Amazon MGM Studios co-producing. Independent in the way most things in Hollywood are independent.

The methodology, branded “Realtime Hybrid Filmmaking,” combines performance capture, virtual production and generative AI across pre-production, shoot and post. The pitch is that actors can see digital environments develop in real time, editorial decisions can be made on the floor, and the gap between idea and final pixel closes. This is a genuine direction of travel in high-end production. It’s also roughly what ILM StageCraft has been doing on The Mandalorian since 2019, minus the generative component.

The generative component is where Luma comes in. Luma CEO Amit Jain says the partnership will bring Luma Agents and the company’s broader model stack into “high quality production.” Luma has been positioning its Ray3 and Uni-1 models as reasoning video systems for professional workflows, and a named production partner gives that pitch a concrete anchor. Whether Innovative Dreams becomes a showcase for Luma’s tech or a captive shop for Wonder Project slate is the question investors should probably ask before Luma’s next round.

The LA jobs framing is the cleverest bit of positioning. Wonder Project CEO Kelly Merryman Hoogstraten calls it “a deliberate bet on Hollywood, on local crews, and on the future of American filmmaking.” At a moment when crew unions are watching AI with understandable alarm, an AI production hub pitched as a jobs programme is either genuinely counter-intuitive or unusually well-advised PR. The release does not quantify how many crew positions the facility will support, or whether generative workflows increase or reduce headcount per production day. Those are the numbers that matter.

For working creators and production professionals, the practical question is whether Innovative Dreams actually opens third-party bookings as promised, and on what terms. The release says the company will provide services to outside studios “across a wide range of genres and budget levels.” If that materialises, it’s a real offer: virtual production plus generative tools on a rentable stage in Manhattan Beach, backed by AWS compute. If it doesn’t, it’s an Amazon in-house facility with a standalone brand.

The Kling integration is the most commercially immediate piece. Kling’s models are widely considered among the strongest currently available, and their arrival inside Firefly means Adobe subscribers can now generate in Kling without leaving Adobe’s environment. Kling 3.0 and 3.0 Omni join Veo 3.1, Runway Gen-4.5, Nano Banana 2 and others in a Firefly lineup that now runs to more than 30 models.

Firefly Video Editor, the browser-based timeline tool, also picks up several upgrades. Enhance Speech, Adobe’s dialogue cleanup feature previously confined to Premiere Pro and Adobe Podcast, is now available inside the Firefly editor. Colour adjustment sliders have been added. And Adobe Stock is now integrated directly into the editor, giving access to more than 800 million licensed assets without switching applications.

The Firefly AI Assistant is the piece that will generate the most commentary but is arriving in the most cautious form. It enters public beta “in the coming weeks” rather than today. In demo form, the assistant takes a conversational prompt, breaks it down, sequences the right tools across Creative Cloud, shows its reasoning, and lets users intervene with traditional sliders and brushes at any point. Adobe is positioning it as agentic creativity with the human in the driver’s seat, which is both marketing language and a reasonable description of the product shape.

Color Mode in Premiere is the quieter but, for working editors, potentially the more significant announcement. Built specifically for editors rather than colourists, it enters public beta today for all Premiere subscribers, with general availability later in 2026. Premiere 26.2 also ships with new Film Impact effects, improved object masking, and a searchable Sequence Index panel. After Effects 26.2 adds AI-powered Object Matte. Frame.io Drive, a desktop app that mounts Frame.io projects as if they were local drives, starts rolling out now.

What this means for AI Video Week readers: Adobe is now the most credible candidate to own the full AI video workflow end-to-end, from generation through editing, colour, audio cleanup and delivery, inside a single subscription stack. For agencies and production companies already paying for Creative Cloud, the case for supplementing it with separate tools is narrowing. Kling inside Firefly is a particular turning point, because Kling-quality generation with Adobe-quality editing and commercial rights in one environment removes several reasons to leave.

The competitive pressure now sits on the pure-play generation platforms. Runway and Pika built their positions on being better at the model than the incumbents. Adobe has now made the incumbent position substantially harder to dislodge. Whether creators continue paying for standalone Kling or Runway subscriptions, or let their Firefly access do the job, is a live question for the rest of 2026.

One note of caution for readers planning workflows around Firefly AI Assistant: Adobe has not confirmed whether it will sit inside existing Firefly credit tiers or require a separate subscription level. Worth watching before committing a team workflow to it.

On the surface, this looks like yet another aggregator wrapping somebody else’s models. The market is already cluttered with those. What makes Shutterstock’s move worth paying attention to is not the models, and not the interface. It is the licensing.

Commercial-ready AI video, with clean indemnification, is still surprisingly hard to buy. Most of the frontier generators ship with carefully hedged terms of service, ambiguity around training data, and very little appetite from platforms to sit behind a customer’s legal team when something goes wrong. Shutterstock already runs an indemnification regime for its stock library, up to $10,000 as standard and $250,000 on enhanced licences, with Premier terms available to enterprise clients. It is extending that architecture to AI-generated output.

For agency producers and brand teams, that matters. It is the difference between an AI video clip you can drop into a paid campaign without anxiety, and one your legal counsel quietly strips out three days before delivery. Shutterstock is betting that a meaningful share of commercial buyers would rather pay a premium for clean provenance than chase the lowest per-second cost across six platforms.

The tool itself is conventional. Text-to-video, image-to-video, the ability to animate existing brand assets or build from Shutterstock’s library. Aspect ratio and duration controls. A model selector that routes prompts to whichever engine the platform thinks fits best. Paul Teall, Shutterstock’s VP of Marketplace Strategy, framed the launch as extending the trust model built up around AI image generation into video.

Pricing sits inside existing Shutterstock subscription structures. The standalone AI image plan starts at $15 a month. Video subscriptions start at $59 on annual billing. The Unlimited plan at $69 a month now bundles AI generation credits alongside the main library. Two free generations per new user act as the acquisition hook.

What this means for AI Video Week readers: if you are producing AI video for clients and have been reassuring them about rights clearance on a case-by-case basis, this is a route that takes that conversation off the table. The output is more expensive than running the same Veo or Runway model direct, but the indemnification is doing real work. For independent creators the calculation is different. The price premium will be harder to justify unless the client is specifically asking for commercial cover.

The more interesting read is competitive. Getty has been circling this territory for a while. Adobe Stock is already woven into Firefly. Shutterstock moving now, with multiple third-party models rather than a proprietary one, says something about how the stock libraries are positioning themselves, as licensing layers rather than generation layers. Whoever owns the indemnification stack for commercial AI video may end up more valuable than whoever builds the models.

The model first appeared as a mystery entrant on the Artificial Analysis Video Arena leaderboard, reaching the number one position in both text-to-video and image-to-video categories before its creator’s identity was disclosed. In the silent video rankings, HappyHorse achieved an Elo score of 1333, opening up a lead of roughly 60 points over Seedance 2.0.

The project was developed by an innovation unit within Alibaba’s newly established Token Hub business group. It was led by Zhang Di, a former vice president at Kuaishou who previously headed the technical development of Kling AI, and who rejoined Alibaba in November after five years away. Zhang’s track record speaks for itself. Kling AI recently reached an annualised revenue run rate of $240 million, making it one of the few AI video platforms to demonstrate genuine commercial traction at scale. That he’s now leading a rival effort underlines how fiercely Chinese tech companies are competing for top talent in generative video.

AI Video Week understands that Kling has new model announcements of its own in the pipeline, so the competitive picture is far from settled. The AI video generation race in China is moving at a pace that makes quarterly product cycles look leisurely.

HappyHorse is a 15-billion parameter open-source model built on a unified self-attention Transformer architecture. It processes text, image, video and audio tokens within a single sequence rather than relying on separate pipelines for each modality. It generates video with synchronised audio natively, supporting lip sync across six languages including English, Chinese, Japanese, Korean, German and French. The model outputs at 1080p and is described as being on track for open-weight release, with an API to follow.

The timing adds extra weight. OpenAI recently discontinued its Sora video generation app, citing a strategic shift towards coding tools and enterprise clients amid high compute costs. Sora was costing roughly $1 million per day to run while its user base had collapsed to fewer than 500,000. Meanwhile, ByteDance was forced to pause the global rollout of Seedance 2.0 following copyright disputes with major Hollywood studios and streaming platforms. The Motion Picture Association sent a cease and desist letter alleging that unauthorised use of intellectual property was systemic rather than incidental.

With Sora gone and Seedance’s international expansion stalled, Alibaba’s entry reshapes the field. An open-source model at this performance level could change the economics of self-hosted generation pipelines for studios and agencies. Native audio with multilingual lip sync, if it works as advertised when the weights ship, addresses one of the persistent technical gaps in AI video production.

The anonymous debut followed a pattern that’s becoming standard among Chinese AI labs. Xiaomi recently used the same approach when its model MiMo-V2 attracted attention under the pseudonym Hunter Alpha. The tactic generates buzz through benchmark performance before the brand reveal. It works, though it raises fair questions about whether leaderboard entries from undisclosed labs carry the same credibility as those from identified ones.

The model is currently in closed beta. An API rollout and open-weight release are expected soon.

The integration, announced this week by HeyGen’s marketing manager Holly Xiao, adds Seedance 2.0’s generation capabilities to HeyGen’s existing avatar platform. Avatar shots now include what the company describes as lifelike movement, consistent likeness, multi-character scenes, and the ability to place AI avatars into any background. The Video Agent tool generates structured videos up to three minutes from a single prompt, including a-roll, motion graphics, and editing. The b-roll generator produces footage from shot descriptions and reference images with per-segment camera control.

For AIVW readers, the more interesting question is what this tells you about the market. HeyGen is the latest platform to build its offering around a model it didn’t create. Seedance 2.0 is ByteDance’s model. HeyGen is wrapping it in workflow tools and selling the result as a production platform. That’s the same aggregation logic that Higgsfield, OpenArt, and others are pursuing, and the same logic that ByteDance is simultaneously undermining by distributing Seedance directly through CapCut.

Whether HeyGen’s avatar and workflow tools add enough value to justify the platform over going direct is a question worth asking. It’s one we explore in depth this week in AI Video Week.

A quick clarification on what “open source” means in practice here, because the term is doing double duty. The model itself is free. Alibaba is releasing the full model weights, architecture, and code under Apache 2.0 licensing, which means anyone can download it, run it on their own hardware, modify it, and use it commercially without paying Alibaba a penny. The open weights for self-hosting are expected within four to eight weeks.

The catch is that running it yourself requires a GPU with at least 24GB of VRAM, plus the technical knowledge to set it up. Most creators don’t have that hardware sitting under their desk. So hosting providers like Together AI run the model on their servers and charge you per clip for the convenience. You’re not paying for the model. You’re paying for someone else’s electricity and infrastructure. At Together AI, that’s roughly $0.10 per video. Other providers like Segmind charge slightly more, around $0.63 per 720p clip and $0.94 per 1080p clip, with differences in speed and features.

That gives creators three routes to the same model. Self-host it for free if you’ve got the hardware. Use an API provider for pennies per clip if you haven’t. Or stick with a subscription platform like Kling or Runway and pay more per clip, but get the interface, the creative tools, and the simplicity that comes with them.

The feature set itself is substantial. Wan 2.7 runs on a 27-billion-parameter Mixture-of-Experts architecture (a design that activates only half its parameters per generation, keeping compute costs manageable) and includes text-to-video and image-to-video at up to 1080p, native audio generation with lip sync, first-and-last-frame control, instruction-based editing of existing clips using plain English, and a 9-grid reference input system for structured composition.

The quality picture is honest rather than breathless. Independent benchmarks place Seedance 2, Kling 3, and PixVerse V6 ahead of Wan 2.7 on raw visual fidelity. Where Wan competes is on the breadth of what it can do in a single pipeline and the cost at which it does it. For high-volume social content, marketing assets, and product video, the output is commercially usable. For hero content where every frame needs to be flawless, the paid platforms retain an edge.

Alibaba has already announced Wan 3.0, targeting 60 billion parameters, 4K resolution, and 30-second generation, expected mid-2026 under the same open licence. The Wan family has over 15,000 stars on GitHub and a growing ecosystem of community adapters and integrations.

Veo 3.1 has been available to developers since October 2025. The Lite tier, launched 31 March, sits at the bottom of an existing three-tier stack and is designed specifically for high-volume applications where per-clip cost is the deciding factor. It’s available now via the Gemini API and Google AI Studio on a paid tier, supporting text-to-video and image-to-video generation at 720p and 1080p resolutions, in both landscape (16:9) and portrait (9:16) formats, with clip durations of four, six, or eight seconds. Pricing is $0.05 per second at 720p. A six-second 1080p clip costs approximately $0.48. A hundred clips costs under $50. For anyone building high-volume video workflows, those are numbers that change the calculation on whether to build in-house or commission from a third-party provider.

The three-tier stack now runs from Lite for volume work, through Veo 3.1 Fast for balanced performance, up to the flagship Veo 3.1 for maximum quality including 4K output. On 7 April, Google will also reduce the price of Veo 3.1 Fast, compressing the cost curve across the entire range simultaneously. Google’s statement ended with a note that further updates are coming, which suggests the pricing move is the opening of a competitive campaign rather than a one-off announcement.

SynthID watermarking is built into Veo 3.1 Lite as standard. Developed by Google DeepMind, SynthID embeds a digital watermark directly into the pixels of generated video, imperceptible to viewers but detectable by software. For creators working to broadcast or platform delivery standards that require AI content labelling, that’s a compliance feature worth knowing about rather than a marketing add-on.

The timing matters. Sora’s consumer app goes dark on 26 April, with the API following on 24 September. Google’s launch, one week after the Sora discontinuation announcement, is a direct play for the developers and production teams now looking for alternatives. Runway, Pika, and Kling are the other obvious candidates, but none currently offers a comparable three-tier API stack at these price points.

Veo 3.1 Lite is an API rather than a finished application — a building block accessed through code, aimed at developers integrating video generation into platforms and pipelines. Creators who want to use Veo’s capabilities without building their own workflow will need to access it through a third-party tool that has already integrated the API. Several have. On the same day as Google’s launch, Higgsfield released Cinema Studio 3.0, a browser-based production environment with camera controls, lighting tools, character consistency and native audio. It’s one of several platforms now running Veo models inside a finished creative interface.