Which models are actually cleared for commercial work

At contract stage, one question decides which model you are allowed to use: if a rights holder makes a claim, who pays. Everything else is a creative preference. Almost nobody publishes a straight answer to that question, so here is the current state of it, with the gaps marked. Practitioner notes, not legal advice.

The most useful public comparison we have found is Everypixel Journal's "Can You Use AI Video Commercially? 10 Model Licenses Compared", published 21 October 2026 and last modified 27 October. It covers licence terms. It does not cover litigation, so read it alongside whatever is happening in court, which is a separate question with separate answers.

Indemnified, with conditions

Google Veo has the cleanest enterprise posture of any frontier model. Veo is named on Google's generative AI indemnified services page, last modified 20 July 2026. Two conditions do the work. Coverage applies only through the Gemini Enterprise Agent Platform API, formerly the Vertex AI API, and only to GA model versions. A preview model, or the same model reached through a consumer app, is not the covered thing. Google also does not assert ownership of output. If a client's procurement team wants one answer they can take to their own counsel without a long conversation, this is the one that survives contact.

Runway indemnifies on the Enterprise plan only. The coverage is wider than most: it addresses claims that the service infringes third-party IP, it extends to other labs' models hosted on Runway, and it sits outside the general liability cap, which is unusual and worth real money. Runway states that it does not restrict commercial use and does not claim ownership of inputs or outputs. Runway's own page on commercial use, published 23 September 2026 and updated 26 October.

The other half of Runway has to be said in the same breath. Standard terms grant Runway broad perpetual rights to inputs and outputs and permit training, with no general opt-out. The no-training commitment applies to Customer Content on Enterprise only. If you are putting a client's own footage into the tool, which is the whole point of the editing models, Enterprise is not an upgrade. It is the minimum.

Adobe Firefly carries copyright indemnity for eligible enterprise customers, for the native Firefly model only. Partner model coverage is separate and limited. Adobe says it plainly on its own page: "Commercial usage eligibility depends on the model used. Partner models may have different usage terms." The operational consequence is the thing most teams miss, and this practitioner piece from 24 August 2026 makes it well: name the specific model in the production spec and the delivery manifest, because "made in Firefly" says nothing at all about clearance. A file can come out of Firefly having been generated by a partner model with entirely different terms.

Luma indemnifies on Enterprise only, with exclusions. Standard terms let Luma use inputs and outputs for model development. API and Enterprise tiers carry no-training commitments. Same shape as Runway: the tier you are on is the term that matters.

MiniMax H3 API offers a limited defence against certain patent and copyright claims, subject to the liability cap. A capped defence is not nothing, but a cap usually resolves to something close to fees paid, which will not cover a pulled campaign.

No indemnification identified

• Seedance and BytePlus.
• Kling consumer terms.
• Wan, both open weights and API.
• LTX open weights.
• Moonvalley.

Kling deserves a note of its own because the output licence reads better than the input licence. Paid members may use, modify and distribute output. But uploads and outputs may be used for model development, and that is revocable only by emailing support, with unconfirmed effect. If a client's unreleased product is in the upload, that is the clause to raise with them before you upload it.

The provenance option

Moonvalley Marey is the model to reach for when a client's legal team cares more about where the training data came from than about where the model sits on a board. Moonvalley says Marey was trained only on licensed footage: "No scraped content. No user submissions. No legal gray zones." Users retain rights, and training is opt-in only. The company raised $84M with CAA, CoreWeave and Comcast Ventures. BetaKit on the round.

“If a studio hands a brand a film that is wholly machine-generated, there may be nothing in it the brand can own in the US.”

Two caveats, both of which we give clients unprompted. We identified no indemnification, so provenance here is a claim about inputs rather than a promise to defend you. And the training-data claims are not independently verified. They are the company's own account. For a rights-sensitive category it is still the obvious pick, because a licensed-footage story is something a client's counsel can understand and repeat, and that has value separate from whether it is the best-looking model available.

Regulation now matters as much as licensing

Clearance is not only a contract question in 2026. Four dates do most of the work.

• 2 August 2026. EU AI Act Article 50 transparency obligations have applied since this date. The duty sits with the deployer, which in advertising means the advertiser and the agency.
• 9 June 2026. New York's synthetic performer disclosure law took effect.
• August 2026. California AB 853 took effect.
• 2 March 2026. The US Supreme Court declined Thaler v. Perlmutter. Purely AI-generated output still receives no US copyright protection. Only the human creative contribution is protectable.

That last one is the commercially important one and it is the one least often raised in a pitch. If a studio hands a brand a film that is wholly machine-generated, there may be nothing in it the brand can own in the US. The brand is paying for an asset that any competitor could, in principle, copy. The answer is not to use less AI. It is to document the human authorship that is already there: the script, the edit, the shot design, the direction of each generation, the grade, the sound design, the selection and sequencing choices. Put it in the deliverable rather than in a folder somewhere. A brand that can point to human authorship in its own archive is in a different position from one that cannot.

The clearance sheet

Everything above collapses into one page that should travel with every job, attached to the production spec at kickoff and to the delivery manifest at handover. Seven fields:

• Model name and exact version, per shot.
• The API route or plan tier it was generated through, since coverage usually attaches to the route and not the model.
• Indemnity: yes or no, and whether it sits inside or outside the liability cap.
• Training status of client-supplied inputs, with the opt-out recorded if one was needed.
• Provenance of every reference image, frame and audio file that went in.
• Disclosure obligations by market of release, not by market of production.
• The record of human authorship, shot by shot.

Nobody will ask you for this sheet on the first job. On the job after the one that went wrong, it is the first thing procurement asks for, and a studio that already has it wins work from studios that do not.