How much does it cost to generate a 5-second AI video? We tested 24 models
We gave 24 video models the same three prompts: a desk scene, a barista handing over a latte, and a product orbit. Hit "Play all" and watch Veo, Seedance, Kling, Wan and the rest attempt the identical shot.
A 5-second AI video clip costs between $0.13 and $1.23 to generate right now, depending on the model, and the spread in what you get back is wider than the spread in price. We ran the same three prompts through all 24 video models in the catalog we use for ad production: one take each, 9:16, 720p, no retries. 72 clips, $38.97 in total, and we logged what every one cost.
Comparison posts about video models usually show each vendor's showreel, which tells you what a model can do on its best day with its own marketing team prompting. This is the opposite: identical briefs, first takes only, including the failures. You can play each grid muted in one go and watch 24 interpretations of the same shot run side by side, which turns out to be the fastest way to understand where this field actually is.
The short version
- A 5-second clip runs $0.13 to $1.23; a usable one costs 3 to 10 takes' worth. Most models cluster between $0.40 and $0.90 per clip.
- Objects are solved, people aren't. The product orbit came back saleable from most of the 24; the barista scene produced a gender-flipped cast, an uncanny selfie arm, and a latte that never left the counter.
- Render speed varies 26x. MiniMax H3 Max returned every clip in 13 seconds; Wan 3.0 Prime took up to 341. If you iterate, speed is a feature you'll pay for in patience.
- Prompt leakage is real in video too. One model filmed its own lighting rig, one printed phantom label text on the bottle, and Veo added a human hand to a scene whose brief said "no people."
- Don't anchor on any single model: Sora is gone, Google retired the whole Veo line from its API on October 22, and April's arena leader fell out of the top 10 by November. Build workflows that can swap models.
How we tested 24 models
Same method as our other cost tests (19 image models, 3 avatar models): one brief, every model, first take kept, priced as charged. Settings were identical wherever a model allowed it: 5 seconds, 9:16, 720p, no generated audio. Where a model forced a different shape, the caption says so; Veo only renders 4, 6 or 8 seconds, so its clips are 4, and Gemini Omni Flash always ships with sound.
The three prompts are three different ad jobs on purpose:
- The desk: a slow push-in on a founder's desk at dusk, with a phone that lights up with a notification halfway through. Tests mood, camera discipline and a timed event. No people allowed.
- The café: a barista slides an iced latte across the counter and smiles. Tests the hard stuff: faces, hands, liquid, object contact.
- The product: a cosmetic pump bottle on a pedestal, orbiting camera, studio lighting, brand colors. The bread-and-butter e-commerce shot.
The lineup is every video model in the production catalog: four from Google, five Seedance tiers from ByteDance, three Klings from Kuaishou, three Grok Imagines from xAI, six from Alibaba (four Wan tiers plus both HappyHorses), MiniMax H3 Max, Black Forest Labs' FLUX 3 Video and Runway's Gen-4.5. Prices are what the catalog charged us per job; it sells generation at cost, so they track the hosts' list rates, with small gaps where billing units differ. Absences worth naming: Sora no longer exists, and Pika, Luma and Hedra aren't in the catalog, so they're not in the test. And everything here is text-to-video. Animating your own product still (image-to-video, which most of these models also offer) is a different test, and the next one we'll run.
Every model, price and render time
Price is what we were charged for one clip (identical across the three runs). Render time is the min-to-max wall-clock range across the three runs, which tells you more than a single number: some models are consistent, others swing 4x depending on the scene.
| Model | Maker | Per clip | Render range | Delivered |
|---|---|---|---|---|
| Veo 3.1 Lite | $0.13 | 24–26 s | 4 s | |
| Seedance 1.5 Pro | ByteDance | $0.14 | 52–58 s | 5 s |
| Grok Imagine Video 1.5 Lite | xAI | $0.16 | 24–37 s | 5.1 s |
| Veo 3.1 Fast | $0.34 | 34–55 s | 4 s | |
| Grok Imagine Video | xAI | $0.37 | 56–62 s | 5.1 s |
| Gemini Omni Flash | $0.41 | 27–32 s | 4 s, with sound | |
| Seedance 2.0 Mini | ByteDance | $0.41 | 98–122 s | 5 s |
| MiniMax H3 Max | MiniMax | $0.43 | 13 s | 5.2 s |
| Wan 2.6 | Alibaba | $0.43 | 35–141 s | 5 s |
| Kling Video v3.0 Standard | Kuaishou | $0.45 | 97–195 s | 5 s, 1080p |
| Seedance 2.0 Fast | ByteDance | $0.49 | 120–194 s | 5 s |
| HappyHorse 1.0 | Alibaba | $0.53 | 79–96 s | 5.2 s |
| HappyHorse 1.1 | Alibaba | $0.53 | 96–116 s | 5.2 s |
| Wan 2.7 | Alibaba | $0.53 | 96–137 s | 5 s |
| Wan 3.0 | Alibaba | $0.53 | 136–166 s | 5 s |
| Kling Video O1 | Kuaishou | $0.60 | 137–202 s | 5 s, 1080p |
| Kling Video v3.0 Pro | Kuaishou | $0.60 | 97–118 s | 5 s, 1080p |
| Gen-4.5 | Runway | $0.64 | 82–223 s | 5 s |
| Grok Imagine Video 1.5 | xAI | $0.74 | 41–47 s | 5 s |
| Wan 3.0 Prime | Alibaba | $0.74 | 80–341 s | 5 s |
| Seedance 2.0 | ByteDance | $0.81 | 164–286 s | 5 s |
| Veo 3.1 | $0.85 | 33–35 s | 4 s | |
| FLUX 3 Video | Black Forest Labs | $0.90 | 65–78 s | 5 s |
| Seedance 2.5 | ByteDance | $1.23 | 142–258 s | 5 s |
MiniMax H3 Max deserves a sentence of its own: 13 seconds per render, three times out of three, on scenes where other models took three to five minutes. When you're burning takes to find a keeper, a model that answers 20x faster changes how you work more than a few Elo points of quality ever will.
All 72 clips, side by side
One tab per prompt. "Play all, muted" starts every clip in view at once (videos pause as you scroll past them and resume when they're back); tap any single clip for controls and sound. Nothing downloads until you press play. The full datasets are downloadable: desk, café, product.
The prompt: Slow push-in on a founder's desk at dusk. A laptop is open on a simple analytics dashboard with one line rising to the right, a ceramic mug in electric blue in the foreground, warm lamp light from the left, soft window light behind. A phone lies face up and lights with a notification halfway through. Shallow depth of field, photographic, calm. No people, no readable text, no logos.
What three prompts exposed
Objects are a solved problem; people are not. The product orbit is the brief where most of the 24 came back saleable: centered bottle, clean light, credible materials. The café scene is where the field's remaining weaknesses live. Seedance 1.5 Pro cast a male barista against a brief that says "she." Wan 2.7's selfie-style reach collapses into uncanny foreshortening. HappyHorse 1.1 left the latte parked on the counter, slide never attempted. FLUX 3 Video fused the cup's lid with the foam into one impossible object. None of these are render-quality problems; they're comprehension problems, and they're exactly what you'll spend retakes on.
Models editorialize. Give 24 systems the words "seamless electric blue backdrop" and several will paint clouds on it (Seedance 2.0 and 2.5, Veo 3.1). My favorite failure: Wan 2.6 filmed the studio itself, softboxes in frame, which would be a lovely behind-the-scenes shot if anyone had asked for one. Kling O1 printed phantom label text on a bottle the brief explicitly called unlabeled. And on the desk scene, whose brief says "no people," Veo 3.1 Fast reached a human hand into frame to tap the phone. The same leakage problem we found in the image model test shows up in motion: instructions are suggestions, and QA has to assume nothing.
The timed event separated tiers. The desk brief asks for a phone that lights up with a notification halfway through, which is a cut-level direction, not a style. Seedance 1.5 Pro nailed it with a visible envelope icon popping onto the screen; Wan 3.0 and Gemini Omni Flash lit the phone convincingly; the rest either ignored the phone or left it glowing from frame one. If your ad needs something to happen at second three, test for that specifically, because most models can't schedule.
The models keep disappearing
Here's the thing nobody making "best AI video generator" listicles wants to admit: the list rots in months, not years.
Sora, the most famous name in the category, no longer exists. OpenAI announced the shutdown on March 24, 2026, closed the app on April 26 and switched off the API on September 24⁵. Google retired the entire Veo line from the Gemini API on October 22, 2026, pointing everyone at Gemini Omni Flash⁶, and fal.ai lists Veo 3.1's removal for November 17¹. The Veo clips in this article were rendered on November 8; by the time you read this they may be unreproducible.
The leaderboards churn just as fast. HappyHorse entered the Artificial Analysis arena pseudonymously in April 2026, took the #1 spot, got revealed as Alibaba's, and by this week's leaderboard it isn't in the top 10⁷; the current top three (Wan 3.0, Utopai X, MiniMax H3) didn't exist a year ago.
The operational lesson: never hard-wire an ad pipeline to one video model. Price your work per clip, keep prompts portable, and treat the model as a line item you re-bid every quarter, because this market re-bids itself whether you participate or not.
The real cost of a usable clip
The per-clip prices above are the honest unit for comparing models and a dishonest unit for budgeting, because first takes mostly aren't keepers. No neutral benchmark exists for yield, so here's the best available data, flagged by source: invideo's production logs report roughly 3 generations per usable shot and a 25% clip-to-final-cut rate, with one 3-minute episode consuming 164 Seedance generations⁸; practitioner write-ups put complex commercial shots nearer one keeper in ten⁹. Call it 3x for simple shots and 10x for hard ones, and a usable 5-second clip costs $1 to $10 in generation.
Against the alternatives that's still absurd. A single licensed HD stock clip runs about $79 on Shutterstock or Adobe Stock¹⁰, and it's the same clip your competitor licensed. Custom b-roll averages around $905 for a 15-second deliverable once a videographer's day rate is amortized¹¹, and arrives in weeks. The honest comparison isn't "AI replaces the shoot." The $1-to-$10 tier now covers the connective footage (product orbits, desk scenes, mood b-roll) that used to be either stock-generic or shoot-expensive. That footage is most of what performance ads actually need. The craft moves to briefs and selection, which is why we publish what full ads cost at five budgets rather than pretending the clip price is the cost.
Which model for which ad job
One take per brief tells you about comprehension, not consistency, so read these as this test plus the volume we run in production, not as a leaderboard:
| Job | First pick | Also good | Why |
|---|---|---|---|
| People on camera | Kling 3.0 Standard ($0.45) | Wan 3.0 ($0.53) | The two most believable baristas: natural smiles, hands that hold things |
| Product shots | Seedance 2.0 Fast ($0.49) | Gemini Omni Flash ($0.41) | Centered, on-palette, materials read true |
| Fast iteration | MiniMax H3 Max ($0.43) | Grok 1.5 Lite ($0.16) | 13 s and ~30 s renders; test ten ideas in the time others render one |
| Cheapest usable | Grok 1.5 Lite ($0.16) | Seedance 1.5 Pro ($0.14) | Sub-$0.20 clips; Grok passed all three briefs, Seedance's one miss was casting the barista male |
| Sound included | Gemini Omni Flash ($0.41) | Seedance 1.5 Pro (~$0.26 with audio) | Omni ships sound whether you ask or not; Seedance renders audio in the same pass (we tested it muted at $0.14) |
| Cinematic finish | Seedance 2.5 ($1.23) | Kling 3.0 Pro ($0.60) | The two with grades you'd mistake for a shoot; you pay for it |
And the judgment call: for ad work, render speed and brief obedience beat arena rank. An ad clip gets moments in a feed before someone decides to scroll (the hook does the work), gets A/B tested against nine siblings, and dies of fatigue in weeks. The model that lets you make 30 obedient attempts this afternoon produces better ads than the model that makes one beautiful clip by dinner. Leaderboards measure the opposite workflow.
Frequently asked questions
What is the best AI video generator in 2026?
There is no single answer, and the blind-preference leaderboards prove it: on Artificial Analysis's text-to-video arena, Wan 3.0, MiniMax H3 and Seedance 2.5 sit within a few Elo points of each other. In our hands-on test, the practical winners split by job: Kling 3.0 Standard and Wan 3.0 handled people best, Seedance 2.0 Fast and Gemini Omni Flash nailed product shots, and MiniMax H3 Max rendered in 13 seconds flat, which matters more than Elo when you iterate.
How much does AI video generation cost?
In our test, a 5-second 720p clip cost between $0.13 (Veo 3.1 Lite) and $1.23 (Seedance 2.5), with most models landing between $0.40 and $0.90. Billing is usually per second of output, so an 8-second clip costs 60% more than a 5-second one. Budget for retakes: practitioner reports and vendor production data put usable-clip yield anywhere from 1-in-3 to 1-in-10 generations, so a keeper realistically costs $1 to $10.
What is the cheapest AI video generator?
Of the models still on sale, Seedance 1.5 Pro ($0.14 for a 5-second 720p clip in our test) and Grok Imagine Video 1.5 Lite ($0.16) are the cheapest we tested that produced usable footage. Veo 3.1 Lite was cheaper still at $0.13 a clip, but Google retired the entire Veo line from the Gemini API on October 22, 2026. Cheap tiers cut corners somewhere: lower resolution ceilings, no sound, or fewer aspect ratios.
How many generations does one usable clip take?
Nobody has published a neutral benchmark, but the honest range is 3 to 10 generations per keeper. Invideo's production logs show about 3 generations per usable shot and a 25% clip-to-final-cut rate, with one 3-minute episode consuming 164 generations. Practitioner write-ups put complex commercial shots closer to one keeper in ten. Simple, reference-locked shots sit at the cheap end; anything with hands, liquids or specific action sits at the expensive end.
Is Sora still available for ads?
No. OpenAI announced Sora's shutdown on March 24, 2026, closed the app on April 26, and turned off the API on September 24, 2026. Its last public rates were $0.10 per second for sora-2 at 720p and $0.30 to $0.50 for sora-2-pro. Every model in our comparison covers that price range from both ends, which is why losing Sora changed less than people expected.
Do AI-generated video ads need to be labeled?
Meta mandates disclosure only for social-issue, electoral and political ads; commercial ads fall under its detection-based AI info labeling. TikTok requires an AI-generated content toggle on ads that contain it. For footage without people, labels rarely matter in practice; for realistic AI humans we add our own on-screen disclosure, because a platform-applied label looks worse than your own.
Sources
- Veo 3.1 family per-second rates and audio pricing split — fal.ai Veo 3.1 (2026)
- Grok Imagine Video rates ($0.020 to $0.080 per second) — xAI model docs (2026)
- Wan 3.0 and 3.0 Prime official per-second pricing — Alibaba Cloud Model Studio docs (2026)
- FLUX 3 Video per-second rates by resolution — Black Forest Labs pricing (2026); Runway Gen-4.5 credit pricing — Runway developer docs (2026)
- Sora discontinuation timeline — OpenAI help center (2026) and Engadget (2026)
- Veo retirement from the Gemini API on October 22, 2026 — Google AI Studio deprecations (2026)
- Text-to-video arena standings — Artificial Analysis Video Arena (as rendered Nov 8, 2026)
- Generations per usable shot in production — invideo production FAQ (2026, vendor-published)
- One-keeper-in-ten practitioner estimate — Techloy (2026)
- Single-clip stock licensing rates — PhotoWorkout stock footage guide (2026) and Adobe Stock plans (2026)
- Average 15-second custom b-roll cost — Advids b-roll pricing survey (2025, small vendor-run sample)