Using AI-Generated Video Commercially: Which Tiers Permit It and What Conditions Apply

A banana boat franchise in Rio de Janeiro submitted a 1080p Kling 2.6 clip of their inflatable raft to Globo's ad clearance team. The network rejected it. The Kling terms of service state no indemnification for copyright claims, and Globo's compliance policy requires an indemnity from the tool provider for any AI-generated material. The franchise owner learned that commercial use of AI-generated video is not a single checkbox. Every platform, every broadcaster, and every client contract may impose requirements beyond what the model's terms grant. What follows catalogues the commercial-use status of each major image-to-video model as of early 2026: which tiers permit it, which terms grant it, and what specific conditions on resale, broadcast, and client work attach. Every claim is attributed to the specific terms document.

Commercial use covers three distinct activities: using generated video in client work, selling the result directly, and monetising on platforms. A model may permit one and prohibit another. Runway Gen-4 is permitted on paid plans only. Luma Ray 2 permits it on all tiers. Stable Video Diffusion prohibits it entirely under the Stability AI Community Licence. The supporting terms document is the only source that matters. Not a marketing page, not a forum post.

Each model below is listed with its specific commercial-use status, the conditions attached, and the document that states them. Models are grouped by licence type: proprietary hosted, open-weights permissive, and open-weights non-commercial.

Proprietary Hosted Models

Veo 3.1 (Google, November 2025): The Vertex AI terms of service state outputs are owned by the user, Google claims no ownership, and the user receives indemnification for copyright claims. Commercial use is permitted. No tier restriction applies. API and VideoFX web app users both get the same rights. The model card labels the licence proprietary hosted only; you do not own the model, but you own the output. Broadcast clearance will still require your own legal review. The indemnification covers Google's training data only, not your use of the clip in an ad.

Kling 2.6 (Kuaishou, January 2026): The Kling terms of service grant the user a perpetual worldwide licence for commercial use of outputs. No indemnification is stated. Commercial use is permitted on all tiers, free and paid. The lack of indemnification is the key gap: if the output contains a copyrighted element that Kuaishou's training ingested, the user bears the risk. The terms do not restrict resale or platform monetisation.

Runway Gen-4 (Runway, February 2026): The Runway terms of service state outputs are owned by the user, Runway claims no ownership, and commercial use is permitted on paid plans only. Free-tier users may not use outputs commercially. No indemnification is stated. This is a direct tier restriction: a solo filmmaker on the free plan cannot sell a clip or use it in client work.

Sora (OpenAI, December 2025): The OpenAI terms of use assign all rights to the user and permit commercial use. No tier restriction is stated. No indemnification is provided. OpenAI's terms explicitly note the user must ensure the output complies with applicable law. EU AI Act transparency obligations are the user's responsibility, not OpenAI's.

Luma Ray 2 (Luma AI, October 2025): The Luma terms of service state outputs are owned by the user, Luma claims no ownership, and commercial use is permitted. No tier restriction. No indemnification. This is one of the cleanest commercial permissions among hosted models: no free-tier block, no resale restriction in the terms.

Pika 2.2 (Pika, December 2025): The Pika terms of service state outputs are owned by the user, Pika claims no ownership, and commercial use is permitted on paid plans only. No indemnification stated. The free plan generates a visible watermark, which makes commercial use impractical even if the legal right existed.

Seedance 1.0 (ByteDance, September 2025): The Jimeng terms of service grant the user a perpetual worldwide licence for commercial use of outputs. No indemnification. Commercial use is permitted on all tiers. The Chinese platform context matters: the EU AI Act and US Copyright Office policies do not apply directly, but any distribution in those jurisdictions remains subject to local law.

Hailuo I2V-01 (MiniMax, August 2025): The Hailuo terms of service state outputs are owned by the user, MiniMax claims no ownership, and commercial use is permitted. No tier restriction. No indemnification. Hailuo is one of the few models that explicitly permits commercial use without a paid-plan gate.

Open-Weights Permissive Terms

Wan 2.1 (Alibaba, July 2025): The Wan model card specifies an Apache 2.0 licence. Commercial use is permitted. The user owns the output. No indemnification is included in the model card. Apache 2.0 grants the right to use, modify, and distribute the software, but the output itself is a separate legal question: the licence covers the code, not the generated video. The same practical risk applies as with hosted models. No indemnity for training-data copyright issues.

LTX-Video 0.9.5 (Lightricks, June 2025): The LTX-Video model card specifies an Apache 2.0 licence. Commercial use is permitted. The user owns the output. No indemnification. This is the clearest open-weights commercial path. Apache 2.0 is a standard permissive licence with no additional restrictions. Run LTX-Video locally on a consumer GPU, generate clips, and sell them or use them in client work without any platform gate.

Open-Weights Non-Commercial Terms

Stable Video Diffusion 1.1 (Stability AI, November 2023): The Stability AI Community Licence explicitly prohibits commercial use of the model and its outputs unless you purchase a separate commercial licence from Stability AI. The model card is clear: non-commercial only. Generate a clip with SVD and sell it or use it in client work, and you are in breach. This is the most restrictive open-weights licence among the models listed here.

Runway Gen-4 and Pika 2.2 both reserve commercial use for paid plans. The practical effect: a free-tier user cannot generate a clip and use it in a YouTube video with monetisation enabled, submit it to a stock footage library, or sell it on a platform. The terms of service for both models state this explicitly. If you are a solo filmmaker or social-media content creator on a budget, factor the paid-plan cost into your project budget before you generate anything. The free tier is for evaluation only.

Kling 2.6, Luma Ray 2, Seedance 1.0, and Hailuo I2V-01 permit commercial use on all tiers. This does not mean the output is free of restrictions. The lack of indemnification is a real risk. But there is no paywall on the right to use the output commercially.

None of the models listed above explicitly prohibit resale of the generated video in their terms of service. The restrictions that do exist come from the platform on which you sell, not from the model provider. Getty Images does not accept AI-generated video at all. Shutterstock has its own contributor guidelines that require disclosure. The EU AI Act transparency obligations mean you must label any AI-generated video you distribute in the EU, regardless of the model's terms.

Broadcast clearance is a separate and more demanding requirement. The banana boat example at the top is not unusual. A broadcaster's compliance team will look at the model provider's indemnification policy. When the provider offers none, which is the case for every model here except Veo 3.1, the broadcaster may require you to indemnify them personally, or refuse the clip outright. This is not a terms-of-service issue. It is a contract law issue between you and the broadcaster. If broadcast clearance is in your path, budget for a legal review of the specific clip and the specific network's policy.

YouTube Creator Studio labels are mandatory for altered or synthetic content that could be mistaken for real footage. The platform's detection system may flag your clip even if the model added its own metadata watermark. YouTube uses its own detection methods. If your clip is flagged and you have not applied the label, the platform may demonetise it or remove it. The same applies to TikTok and Instagram. Both have synthetic-content disclosure policies.

The metadata watermark versus visible watermark distinction matters. Pika 2.2's free tier adds a visible watermark. Sora and Veo 3.1 embed metadata watermarks in the file header. No platform guarantees it will read that specific metadata format. The safest approach for platform monetisation: apply the platform's own disclosure label manually, regardless of what the model did.

The contrast between LTX-Video's Apache 2.0 and Stable Video Diffusion's non-commercial Community Licence is the most important distinction in the open-weights space. Apache 2.0 grants you the right to use the software for any purpose, including commercial. The Stability AI Community Licence explicitly withholds that right. If you are an open-source tinkerer running models locally on consumer GPUs and you intend to sell the output, LTX-Video and Wan 2.1 are your options. SVD is not.

Wan 2.1 uses Apache 2.0 as well. Its native resolution tops out at 720p and it requires a more powerful GPU than LTX-Video. The trade-off is resolution versus hardware cost. For a solo filmmaker generating b-roll, LTX-Video's 480p native resolution may be too low for broadcast use. Upscaling with FFmpeg is a standard post-processing step that adds time but can produce acceptable web-quality results.

Only Veo 3.1 provides indemnification for copyright claims. Every other model explicitly states that no indemnification is provided. This is not a minor detail. If a copyright holder claims your generated clip infringes their work, you bear the entire legal and financial risk. The model provider will not defend you.

For a legal or compliance person clearing AI-generated video for broadcast or advertising, the absence of indemnification is the single most important fact here. Your organisation must conduct its own copyright review of the output, or accept the risk. The US Copyright Office's human-authorship requirement adds another layer: the office will not register a copyright for a fully AI-generated work. You cannot sue for infringement of that work either. Your commercial value lies in the right to use the output, not in copyright ownership of it.

If you are a filmmaker or content creator using AI-generated video in client work, your contract with the client should address three things the model's terms of service do not cover. First, the client needs to know that no indemnification is provided by the model provider. If a copyright claim arises, you and the client share the risk. Second, the client should explicitly agree to accept the output under the model's commercial-use terms, which may restrict resale or broadcast. Third, the contract should specify whether the client can use the output on platforms with monetisation enabled. The platform's own policies may impose additional restrictions.

The terms of service for Runway Gen-4, Pika 2.2, and all other models listed here are the controlling documents. Your contract cannot override them. It can allocate the risk between you and the client. A standard practice is to include a clause stating the AI-generated video is licensed under the model's terms, and the client accepts those terms as part of the project agreement.

The most common failure is assuming that because a model permits commercial use, all downstream uses are covered. They are not. A clip that is fine for a YouTube video may be rejected by a broadcaster because the model provider offers no indemnification. A clip that is fine for a client presentation may violate the platform's monetisation policy because the metadata watermark is not recognised by the platform's detection system. The model's terms of service, the platform's policies, and the client or broadcaster's requirements are separate documents. All must be satisfied. Read them before you generate, not after.