A 30-second AI video spot costs between about $19 and $86 in API fees once you pay for the takes you throw away, and the model fee is the smallest risk in the budget. Every vendor bills per generated second, so a shot you reject costs exactly what a shot you keep costs. Our recommendation: run published brand work on Veo 3.1 through Google Cloud, because it is the only model here that is generally available, on a published generated-output indemnity list, and marked with both C2PA and SynthID. Use Runway's API as the creative team's sandbox. Do not plan anything around Sora: OpenAI shut the API down on September 24, 2026.
All prices below were checked on each vendor's live page on October 31, 2026.
| Option | Billed rate (1080p unless noted) | Indemnity for outputs | Provenance | Who should not pick it |
|---|---|---|---|---|
| Veo 3.1 on Google Cloud (Agent Platform, formerly Vertex AI) | Google's console rate card; resold by Runway at $0.40/s with audio, $0.20/s without | Yes, GA versions only | C2PA and SynthID | Teams outside the US that need in-region generation |
| Veo 3.1 Fast | $0.15/s with audio, $0.10/s without (via Runway) | Yes on Google Cloud, GA versions | C2PA and SynthID | Anyone who needs 4K output |
| Gemini Omni Flash (Gemini API) | about $0.10/s at 720p | Confirm with Google before relying on it | Not documented in sources we found | Teams that need native 1080p; higher resolutions are upscaled |
| Runway Gen-4.5 (Runway API) | $0.12/s | None in public terms | Not documented in sources we found | Legal teams that require a vendor defence |
| Luma Ray3.2 (Luma API) | $1.20 per 5s clip, $3.60 per 10s clip | None in public terms | Not documented in sources we found | Brand work that needs any cover at all |
| Adobe Firefly Video (enterprise) | Contact sales | Select Firefly outputs, enterprise only | Content Credentials | Teams that will not sign an Adobe enterprise agreement |
| OpenAI Sora 2 / Sora 2 Pro | Shut down September 24, 2026 | n/a | n/a | Everyone |
What Each Video API Actually Charges per Second
Every live option bills per second of generated video, and the spread between the cheapest and most expensive 1080p rate is roughly four times. The cleanest public price list belongs to Runway, because its API resells other labs' models next to its own. Runway's API pricing page sets one credit at $0.01 and lists Gen-4.5 at 12 credits a second ($0.12), Gen-4 Turbo at 5 ($0.05), Veo 3.1 at 40 credits with audio and 20 without ($0.40 and $0.20), and Veo 3.1 Fast at 15 and 10 ($0.15 and $0.10). The same page lists Kling, Seedance, Wan and Hailuo models from $0.02 to $1.50 a second depending on resolution, so a single Runway key gives a creative team most of the market on one invoice.
Google sells Veo 3.1 directly through Google Cloud. Its Veo 3.1 model page lists veo-3.1-generate-001 and veo-3.1-fast-generate-001 as generally available and veo-3.1-lite-generate-001 as preview. Google's Cloud pricing page did not render a readable rate table when we checked, so take your Veo rate from the Google Cloud console or your account team rather than from the third-party "Veo pricing" guides, which disagree with each other by more than double. Runway's resale price is the floor of what you can verify in public today.
The Gemini API is a separate storefront, and it changed this month. Google's Gemini API video docs say the Veo 3.1 preview models there were shut down on October 22, 2026, and point developers to Gemini Omni Flash. The Gemini API pricing page bills Omni Flash video output at $17.50 per million tokens, at 5,792 tokens per second of 720p video, which works out to about $0.10 a second. Google's changelog says the model went GA on August 27, 2026 and that 1080p and 4K exports are produced by upscaling, not native generation.
Luma prices per clip, not per second. Its API pricing page lists Ray3.2 text or image to video at $1.20 for a 5-second 1080p clip and $3.60 for a 10-second one, or $0.24 and $0.36 a second. The 10-second clip costs three times the 5-second one, so longer takes cost more per second. HDR output is billed at twice the SDR price.
Adobe does not publish an API price for Firefly video. The Firefly plans page sells seats: $9.99 a month buys 2,000 credits and up to 20 five-second videos, and $199.99 a month buys unlimited access to the Firefly Video Model in the Generate Video tool. That is a designer's tool budget. The API and the indemnity come through an enterprise contract, and the price is a sales conversation.
The Sora API Is Gone
OpenAI's Sora 2 API no longer exists, so any budget or vendor shortlist that still includes it is out of date. OpenAI's Sora 2 model page says the models and the Videos API "were shut down on September 24, 2026 and are no longer available," and that "there is no one-to-one replacement API." Its deprecations page shows developers were notified on March 24, 2026, and lists the Videos API and five model IDs, from sora-2 to sora-2-pro-2025-10-06, with no recommended replacement for any of them.
The last published rates are still on the model page: Sora 2 at $0.10 a second at 720p, and Sora 2 Pro at up to $0.70 a second at 1080p. A team that built a production pipeline on Sora in early 2026 got six months' notice and nowhere to migrate inside OpenAI. We covered the shutdown's business side in our April piece on OpenAI's Sora decision.
A Worked Budget for a 30-Second Spot
At four takes per shot, a 30-second spot costs $19 to $86 in model fees, and you pay for 6.4 seconds of video for every second that airs. Here is the workload, stated so you can change it:
- Final cut: 30 seconds, six shots of 5 seconds each.
- Clip length: 8 seconds per take, so the editor has room to trim. Veo 3.1 generates 4, 6 or 8 second clips; Luma is priced on its 10-second clip.
- Retake ratio: four takes per shot. That comes from one of the few published production logs: on a 3-minute animated episode, 164 clips were generated and 41 made the final cut, a 25% selection rate, and the editor used an average of 5 seconds from each 15-second clip. The source is a video-agent vendor with an interest in the topic, so treat it as one data point.
- Generated seconds: 6 shots × 4 takes × 8 seconds = 192 seconds.
| Model | Rate | 4 takes per shot (192s) | 10 takes per shot (480s) |
|---|---|---|---|
| Veo 3.1 with audio (Runway resale) | $0.40/s | $76.80 | $192.00 |
| Veo 3.1 without audio | $0.20/s | $38.40 | $96.00 |
| Veo 3.1 Fast with audio | $0.15/s | $28.80 | $72.00 |
| Gemini Omni Flash 1.1 at 1080p (Runway resale) | $0.15/s | $28.80 | $72.00 |
| Runway Gen-4.5 | $0.12/s | $23.04 | $57.60 |
| Gemini Omni Flash at 720p (Gemini API) | about $0.10/s | $19.20 | $48.00 |
| Luma Ray3.2, 10s clips at 1080p | $3.60/clip | $86.40 (24 clips) | $216.00 (60 clips) |
Two things follow. First, the model fee is small. Even a messy production at ten takes per shot on the most expensive rate stays under $250, which is less than a few hours of an editor's time. Second, the retake ratio multiplies everything, and it is the number nobody puts in the budget. Character consistency is where it climbs: the same production log reports that around 40% of final shots were stitched from two or more generations. If your spot has a recurring person or product, budget the 10-takes column.
The larger costs are review hours spent on rejected takes, a quota that throttles you on deadline day, and a model retirement that forces you to regenerate a campaign's look on a different model.
Who Owns the Output, and Who Defends You
Google is the only vendor here that publishes a written indemnity naming its video model; Runway and Luma put the risk on you. For a spot that runs under your brand, that outweighs a few cents a second.
Google's generative AI indemnified services list (April 22, 2026 snapshot) names the Gemini Enterprise Agent Platform API, formerly the Vertex AI API, used with generally available versions of models including Veo. The service specific terms say pre-GA offerings "are not covered by any SLA or Google indemnity." That puts Veo 3.1 Lite, still in preview, outside the cover, while Veo 3.1 and Veo 3.1 Fast are inside it. As we found for images in our AI image indemnity comparison, Google's cover excludes trademark claims from use in commerce and applies to unmodified output, so archive the raw generation, its prompt and its timestamp before an editor touches it. If you buy Gemini Omni Flash through the Gemini API, get written confirmation that your endpoint is on the list.
Runway's terms of use say the company "does not claim ownership of any of your Inputs or Outputs" and does not restrict commercial use. They also say "your use of outputs is at your sole risk," require you to indemnify Runway, and grant Runway a perpetual licence to use inputs and outputs "to train and improve its AI models." If you send unreleased product footage as a reference image, that is the licence it falls under unless your contract says otherwise. Ask for the API or enterprise agreement in writing before a brand team uploads anything confidential.
Luma's terms of service give you ownership of outputs "to the greatest extent permitted by law," allow commercial use only for outputs produced during an active paid subscription, offer no indemnity, and require you to defend Luma against claims arising from your inputs or outputs. Luma may also use paid-plan inputs and outputs to train its models. Luma says API access has its own terms, which we could not review, so check them for a training opt-out.
Adobe's plans page says "eligible enterprise customers can also receive IP indemnification for select Adobe Firefly outputs." Its partner-model terms exclude outputs from non-Adobe models inside Firefly from that cover, apart from partner models Adobe lists as eligible for its supplemental coverage. A designer who switches the Firefly model dropdown to an unlisted partner video model can leave the indemnity without noticing.
Provenance and Disclosure: C2PA, SynthID and the EU AI Act
If your video reaches EU audiences, both the model provider and your team now carry disclosure duties, and Google's stack is the one that meets the provider side by default. Google DeepMind says videos made with Veo are marked with SynthID, an invisible watermark, and Google's Veo 3.1 model page lists Content Credentials (C2PA) as supported on all three Veo 3.1 models. C2PA is a signed metadata record attached to the file that says which tool made or edited it. It survives your edit suite only if your export keeps it, and many social platforms strip it on upload. We did not find Runway or Luma documentation on C2PA for API outputs; ask both.
The law splits the work. EU AI Act Article 50 requires providers of systems that generate synthetic video to mark outputs "in a machine-readable format and detectable as artificially generated," and requires deployers of deep fakes to disclose that the content was generated. For evidently artistic, creative or fictional work, the deployer duty narrows to disclosure "in an appropriate manner that does not hamper the display or enjoyment of the work." Article 50 applies from August 2, 2026. The AI Omnibus gave the provider marking duty a grace period to December 2, 2026 for systems already on the market, with none for systems launched after August 2.
In practice your team owns the disclosure on a realistic ad that shows a person, whichever vendor you use, because Article 50(4) puts it on the deployer.
Quotas and Regions That Cap Production Volume
Veo 3.1 runs in one US region with short clips, and Runway caps new accounts at $100 a month, so check both before you schedule a deadline. Google's Veo 3.1 model page lists a single region, us-central1, for all three models; a quota of 50 regional online prediction requests per minute per base model; at most 4 output videos per prompt; and 4K output only on the standard model, not Fast or Lite. A European team with a data-residency requirement has to clear sending prompts and reference images to the US before it starts.
The same page lists Veo 3.1's discontinuation date as "November 17, 2026 or later." Google's dates are the earliest a model can retire, not a promise that it will, but it means the GA model you approve this quarter may need a successor before your next campaign. Keep the reference images that define a recurring campaign's look in your own asset system, so you can regenerate it on a successor model. We compared vendors' notice windows in our model deprecation policy guide.
Google's Responsible AI for Veo page also says requests that generate children are rejected unless personGeneration is set to allow_all or the project is on an allowlist, and photorealistic depictions of prominent people are rejected unless the project is allowlisted. A family-product spot needs that setting or allowlist approval before production week.
Runway's usage tiers start at 1 to 2 concurrent jobs and a $100 monthly spend cap. Tier 2 opens after a $50 purchase, and Tier 5 allows 20 concurrent jobs and $100,000 a month after $5,000 in purchases. All video models share one concurrency limit per project, and exceeding the rolling 24-hour generation cap returns HTTP 429. A team that signs up the week of the shoot will hit Tier 1 limits on its first afternoon of retakes.
Which Option Loses, and Who Should Pick Each
Luma Ray3.2 is the weakest choice for enterprise brand work: at $0.36 a second for a 10-second 1080p clip it costs nearly as much as Veo 3.1 with audio, and it comes with no indemnity, commercial rights only while you keep paying, and a training licence over your paid outputs. Its 5-second clips are cheaper per second, and it has a place for internal concept work. For a campaign that legal has to sign off, pick one of the options below.
Pick Veo 3.1 on Google Cloud for anything published under your brand. It is GA, on Google's indemnified list, and marked with C2PA and SynthID. Do not pick it if you cannot send data to us-central1, or if you need the Lite tier's price, which is preview and outside the cover.
Pick Runway's API for the creative team's exploration and for animatics, where seeing ten models on one invoice is worth more than a defence clause. Do not pick it for final assets in a regulated industry unless your enterprise contract adds the indemnity and the training opt-out the public terms lack.
Pick Adobe Firefly if your designers already work in Adobe enterprise and you want the cover inside their tool. Do not pick it if you will not sign an enterprise agreement; seat plans carry no indemnity.
Pick Gemini Omni Flash for cheap 720p drafts and conversational edits. Do not pick it for a 1080p or 4K master until you have checked the upscaled output and confirmed your Gemini API endpoint's indemnity status. Our Runway and Google Vertex AI tool pages cover the platforms themselves.
What Predicts Regret on This Purchase
The buyers who regret this purchase chose on per-second price and skipped the questions that decide whether the asset survives legal review. The criteria that actually predict regret:
- Is the exact model version GA and on the vendor's indemnity list? Preview models at Google and unlisted partner models at Adobe fall outside cover.
- Does the vendor train on your inputs? Reference images of an unreleased product are the most sensitive thing your team will upload.
- What is the retirement date, and where do you migrate? Sora gave six months and no replacement. Veo 3.1 lists November 17, 2026 or later.
- Can you get the volume on deadline day? Region, per-minute quota, concurrency tier and spend cap.
The answer changes if your work never ships externally (then Runway's breadth or Gemini Omni Flash's price wins outright), if you are an EU-only team with strict residency (then Veo's single US region may rule it out), or if your designers live in Adobe (then Firefly's cover inside the tool beats Google's cover in a pipeline they will not use).
This Week: Ask your Google account team for the Veo 3.1 and Veo 3.1 Fast rates on your contract, and confirm in writing which Veo and Gemini endpoints your indemnity covers.
This Month: Run one real shot through Veo 3.1, Gen-4.5 and Gemini Omni Flash, count the takes until the creative lead approves one, and use that ratio in your budget instead of ours.
Before the Next Campaign Brief: Set a policy that names which models may produce published assets, require the raw generation and prompt to be archived, and add an AI disclosure line to your EU ad templates.
Continue Reading
- OpenAI Kills Sora After $200M Loss: Enterprise-Only Pivot
- Adobe, Google and Getty Offer Different AI Image Indemnities
- Model Deprecation: Bedrock Keeps Claude 4 Months Past Anthropic
- Meta's Commercial Terms Put All AI Ad Risks on Advertisers
- HappyHorse-1.0: The #1 AI Video Model Goes Enterprise
- Adobe's Topaz Pledge Names Apps and Models but Never the API