Synthetic film from managed services: prices, terms and how to choose

What you trade

Part one ended on the cost of the open-weights way: tens of gigabytes of weights per stage, on the defaults of its calculator 127 GPU hours on an L4 for one finished minute, a licence matrix to review, and a person watching every generation. The managed services sell the same outputs per second, per character or per seat. The GPUs and much of the quality work become theirs. What stays yours is the brief, the review, and the terms of what you generate. What you give up is control: over the model, its version, and whether it still exists next month.

A month of churn

That last point is not hypothetical. On 24 September 2026, OpenAI shut down the Sora 2 models and its Videos API, with no one-to-one replacement. On 22 October 2026, the Veo 3.1 preview models were shut down on the Gemini API, where Google now points to Gemini Omni Flash; the generally available Veo 3.1 models remain on Google Cloud’s agent platform. Runway retired Gen-4 Aleph on 30 July 2026. A pipeline built on a service has to be built to change service.

Generated shots

Service	Model	Public price	Clip length	Notes
Gemini API	Gemini Omni Flash 1.1	$17.50 per million video output tokens, about $0.10 per second of 720p	3 to 10 s, extendable to 40 s	Audio generated by default; SynthID watermark on every video.
Runway API	Gen-4.5	12 credits per second at $0.01: $0.12/s	2 to 10 s	Gen-4 Turbo $0.05/s; Runway also resells Veo 3.1, Kling 3.0 and Omni Flash.
Runway API, resale	Kling 3.0	$0.09 to $0.17/s depending on tier and audio; $0.43/s in 4K	up to 15 s	Kling’s own API prices could only be read from search snippets.
Luma API	Ray3.2	$0.30 per 5 s clip at 720p, $0.06/s	5 or 10 s	The page’s 10 s price is three times the 5 s price, not twice.
MiniMax API	MiniMax H3	$0.08/s at 768p, $0.13/s at 2K	4 to 15 s	The same model’s open weights are not licensed in the EU or the US; the API is a separate contract.
BytePlus ModelArk	Seedance 2.5	$10.70 per million tokens; the page estimates $0.231/s at 720p	not stated	Failed generations are not billed.

Voice

Service	Public price	Commercial use	Notes
ElevenLabs API	$0.08 per 1,000 characters (Multilingual v2, v3); $0.04 for Flash and Turbo	All paid plans, outside beta services	Plans from $6 a month; the free plan has no commercial licence.
Cartesia	Pro $5 a month for 100,000 characters; overage $0.038 to $0.065 per 1,000	From Pro	Training on content allowed, with an opt-out form.

At an assumed narration pace of 150 words a minute, about 900 characters, that is about $0.07 a minute through the ElevenLabs API, before retakes.

Presenters

Service	Public price	Included	Notes
Synthesia	Starter $29 a month ($18 billed yearly); Pro $89 ($64)	About 12 and 60 minutes of video a month	Does not use customer inputs or outputs to pre-train its models; the customer owns the output.
HeyGen	Creator $29 a month; Pro $49; Business $149	Credits per plan; API from $0.0167 to $0.0667 per second of output	The free plan forbids commercial use; the terms allow training on your content, with no opt-out described.

Music

Service	Public price	Commercial use	Notes
ElevenLabs Music	API $0.15 per minute	Self-serve plans: all uses except film, TV, radio and games	3 s to 5 min per generation; attribution required on the free plan.
Suno	Pro and Premier plans	Paid plans only; the user owns the songs	No public self-serve API; the terms give Suno a perpetual licence to use content for training.
Udio	Not readable	Not readable	Downloads disabled since its partnership with Universal Music Group of 29 October 2025.

Editing

Service	Public price	What it adds	Notes
Descript	From $16 to $65 per person and month, by plan and billing	Editing a video by editing its transcript; 10 to 40 media hours and 400 to 1,500 AI credits a month on paid plans	The terms allow training on inputs and outputs unless you opt out; the pricing page lists the opt-out under Enterprise only.
Adobe Firefly	Credit plans, from 2,000 to 50,000 credits a month; prices vary by country	About 100 credits per five-second generated clip; partner video models from Runway, Kling and Luma, audio from ElevenLabs	Adobe’s own models are trained on licensed and public-domain content, and Adobe does not train on subscribers’ content; outputs of partner models are your responsibility.

Rights and training

The price is the easy part to compare. The terms differ more: who owns the output, whether your scripts and footage train the vendor’s next model, and whether the free tier can be used for work at all. Of the services above, Synthesia states it does not train on customer content, ElevenLabs offers an opt-out, Google’s paid Gemini API tier does not use prompts or responses to improve its products, and HeyGen and Suno grant themselves training rights with no opt-out described. For a company, those lines matter more than a cent per second.

A month, both ways

The calculator compares one month of production. Its self-hosted GPU hours default to part one’s estimate for a 14B model on an L4, and you can replace them with your own.

Figure 1 · A month of production, both ways

Finished minutes per month 10 Share of the film in generated shots 30 % Generations per kept shot 4 Managed video, $ per generated second Managed voice, $ per finished minute Presenter plan, $ per month Self-hosted GPU hours per generated minute, at one generation GPU price, $ per hour Engineering hours per month to run a stack Cost of an engineering hour, $

Managed, per month

$162

Self-hosted GPU

$389

Self-hosted people

$3,200

GPU hours, self-hosted

381.6 h

Managed = minutes × share × 60 × generations × price per second + minutes × voice price + presenter plan. Self-hosted = minutes × share × GPU hours per generated minute × generations × GPU price + engineering hours × rate. Defaults: Gemini Omni Flash at about $0.10/s; ElevenLabs at $0.08 per 1,000 characters for an assumed 900 characters a minute; Synthesia Pro at $89; part one’s 31.8 GPU hours per minute for a 14B model at 720p on an L4, at 60 % utilisation; the g6.xlarge price in Paris, $1.0216 an hour. At the 33 % utilisation measured in part one, the self-hosted figure rises to 57.5 GPU hours per generated minute. Engineering hours and their cost are your assumptions. Review time is the same both ways and is left out, as are the other self-hosted stages.

A workflow with services

Each stage of the pipeline in part one has a managed counterpart. The table maps one to the other; what the services remove is the GPUs, the weights and the integration of each model.

Stage in part one	Managed option	What to watch
Script and shot list	Any language model, by chat or API	The prompts still have to match what each service can do.
Keyframes and shots	Gemini Omni Flash, Runway, Luma, MiniMax, Seedance	Priced per second; clips of 2 to 15 seconds; models are renamed, replaced or closed within months.
Presenter and lip-sync	Synthesia, HeyGen	A real person’s face or voice needs that person’s consent; HeyGen’s free plan excludes commercial use.
Voice	ElevenLabs, Cartesia	Priced per character; acronyms and names still need checking by ear.
Captions	ElevenLabs Speech to Text ($0.22 an hour), Descript	Proper nouns and line breaks.
Score	ElevenLabs Music ($0.15 a minute by API)	Self-serve plans exclude film, TV, radio and games.
Compositing, mix, encode	Descript, Premiere Pro with Firefly	The edit remains a person’s work.

What does not change: the brief, the choice between generations, the review of every licence and every term, and the edit itself. The second way buys back the infrastructure, not the judgement.

Choosing a way

Figure 2 · Four questions

More than about 30 finished minutes a month Scripts or footage may not leave your infrastructure A team already runs GPU inference in production You need a model of your own, trained on your material

The second way. Without a team that already runs GPU inference, the open stack is a new product to build and maintain. Buy the services, and spend the effort on the brief and the review.

A rule of thumb drawn from the arithmetic of both parts, not a procurement method.

What this part does not establish

Prices and terms are those on the vendors’ pages on 25 October 2026. Some pages could not be read in full; the table says so where it matters.
The calculator is arithmetic on public prices and your assumptions, not a quote.
No service was tested for this article, and the quality of their outputs is not compared.
Some prices could only be read in search results, such as Google Cloud’s Veo pages and Kling’s own API page; they are left out, or given through a reseller’s page.

Sources

Video: Gemini API pricing, Veo on the Gemini API, OpenAI deprecations, Runway API pricing and changelog, Luma API pricing, MiniMax pricing, BytePlus ModelArk pricing.
Voice and presenters: ElevenLabs API pricing, ElevenLabs terms, Cartesia pricing, Synthesia pricing, Synthesia AI governance, HeyGen pricing, HeyGen terms.
Music: Eleven Music terms, Suno pricing, Suno terms, Udio and UMG.
Rights and training on the Gemini API: Gemini API terms.
Editing: Descript pricing, Descript terms, Adobe Firefly, Adobe generative credits.