01 / ROUTE
Use cheap drafts, promote survivors.
Draft with Wan, Lite, Fast, or Mini tiers. Spend Veo / Kling Pro money only on the 10–30% of concepts worth a final render.
Research brief + live unit-economics model
AI GENERATION APIs.
COMPARED CLEARLY.
Pick a use case, compare like-for-like costs, then model the number of usable videos or images your product can safely include per user.
01 · Start here
Shortlist the right route in one table.
These are practical starting points, not universal winners. “100 usable” includes 2.5 attempts per keeper so a cheap but unreliable route does not look artificially attractive.
| Model + provider | Best starting use | Configuration | One output | 100 usable outputs | Quality benchmark | Diligence | Next step |
|---|
02 · What the research says
The headline rate is only half the bill.
A $0.05 clip that takes four attempts costs more than a $0.15 clip that works first time. Normalize for keeper rate, audio, resolution, duration, platform fees, and delivery—not just “price per generation.”
02 / RETRIES
Budget 2–4 attempts per keeper.
Retry multiplier dominates unit economics. Track usable-output rate by model, prompt class, and user cohort.
03 / RAILS
No aggregator wins every route.
Pika is often cheapest versus fal/Runway; WaveSpeed and Atlas lead some budget routes; direct Google matches or beats aggregators on Veo.
04 / RISK
Cheap can hide provenance risk.
Very low rates deserve a terms, privacy, model-authority, moderation, retention, and capacity review before production traffic.
03 · Interactive cost lab
Model the cap before users find it.
Set your route, user mix, output volume, retries, and subscription target. The calculator returns monthly cost, gross margin, and a safe monthly cap for each user who generates.
—
—
Monthly COGS
$0
Generation + delivery + app overhead
COGS / active generator
$0
Not averaged across inactive users
Gross margin
0%
At the entered plan price
Safe cap / generator
0
Usable keepers at target margin
Cost stack
Model inference$0
Storage + egress$0
App + moderation overhead$0
Payment processing$0
Provider fixed fee$0
Total monthly COGS$0
Scale curve · same behavior
Recommendation: —
04 · Full model matrix
Same unit. Same duration. Fewer surprises.
Rates below are public list prices, before tax and negotiated volume discounts. “5s cost” normalizes per-second routes; fixed-duration models keep their published request price.
0 routes
| Model / creator | Provider rail | Use case | Configuration | List rate | 5s / asset cost | 100 usable outputs | Quality Elo | Prod. diligence | Source |
|---|
05 · Provider rails
Aggregator, direct API, or creative suite?
They solve different problems. API rails optimize integration and unit price. Creative suites add workflows and identity tools. Direct APIs reduce middlemen but increase vendor integration work.
06 · Asset quality + examples
Quality needs a benchmark and a gallery.
Elo scores come from blind, crowdsourced pairwise preferences on Artificial Analysis. They are a useful prior—not a substitute for your own prompt set, brand constraints, and keeper-rate test.
Job to be done
Best starting family
Budget lane
What to test
Cinematic hero clip
Veo 3.1 / Kling 3 ProRealistic physics vs. stylized direction.
Veo Fast / Kling StdDraft silent when audio can be added later.
Faces, hands, text, camera path, audio sync, first/last frame compliance.
High-volume ads
Seedance 2.0Strong prompt and reference adherence.
Seedance Fast / WanGenerate variants, promote winners.
Product geometry, spokesperson identity, usable rate at 3 variants per brief.
Motion transfer
Kling Motion ControlExplicit driving-video or preset motion.
Wan Animate / Act TwoLower cost for avatar-like motion.
Limb fidelity, occlusion, drift, input-video billing, retained audio.
Images + edits
GPT Image 2 / Nano Banana 2Highest preference scores.
Seedream / Flux KleinStrong price-performance for volume.
Typography, brand color, reference identity, edit locality, 1K→4K price step.
07 · Sources + method
Every number should be re-checkable.
Pricing changes quickly. Links go to the provider pricing/docs or the independent benchmark. The “source verified” label means a public page exposed the exact rate used here.
Method. Prices were researched on 11 August 2026 and normalized to USD per generated second, per fixed request, or per image. Five-second costs are mathematical comparisons, not claims that every endpoint accepts a five-second duration. BytePlus direct rates use its official five-second, no-video-input examples; actual Seedance billing is token-based. Quality Elo is copied from the current Artificial Analysis text-to-video-with-audio or text-to-image leaderboard and is only compared within the same modality. Taxes, failed-call policy, volume discounts, reference-input charges, minimum durations, moderation rejects, storage retention, territory restrictions, and provider-specific rounding can change realized cost. Run a 50–100 prompt benchmark and inspect provider terms before production commitment.