Best AI video generator for ads in 2026: cost per approved variant, not a beauty contest

The cost pillar across every key is how much AI video generation costs; the 15-second cost walkthrough this page borrows is AI UGC ads cost. Same-prompt footage is a compare of its own, still to come. lotgen.ai quotes twelve generation keys. It is not an avatar / UGC-actor tool.

The short answer

How to read this table

• Dollars and seconds only. 1 / 3 / 5 attempts are a modeling assumption, not a measured keep rate. There is no lotgen.ai observed keep rate on this page.
• Planning length is 15 seconds, audio on / default, one aspect, one prompt. If the API will not bill 15 seconds, the 1 / 3 / 5 cells use the closest official billed length and the max-length column says so.
• Google / Veo dollars are Gemini API, paid tier, audio default. Vertex prints a silent column that this table does not mix in. Veo 3.1 lengths are 4, 6, 8 seconds only.
• fal and evolink are their own checkout. Matching Kling’s 720p audio row on fal on the day checked is a coincidence, not a promise that 1080p or 4K match.
• BytePlus per-second is computed from official list 5.6 USD / M tokens and the published formula, plus the closing frame ARK bills (measured by lotgen). See the cost section.
• Alibaba dollars are Model Studio International wan2.7-t2v ($0.10/s 720P, $0.15/s 1080P). Live lotgen /pricing still prints Wan 2.7 1080p as unpriced. That is the product catalog, not Alibaba list. Do not paste evolink Wan as Alibaba list.
• MiniMax dollars are official PAYG list, not the 5% video-package inference.
• Runway’s gen4.5 pricing row prints no cap and no ratio; the length and ratio cells are lotgen’s own test of the route (2026-09-06), labeled as such.
• 9:16 and 1:1 are from that provider’s docs on the day checked. “Not printed” means this page does not invent them. Meta Feed still accepts 9:16 and may display 1:1; that is an ads-spec fact, not a generator capability.

Meta’s Facebook Feed video specs on 2026-09-18: duration 1 second to 241 minutes, so 15 seconds is inside that window. Recommended ratio 4:5 (1440 × 1800). Tooltip: ratios from 16:9 to 9:16 are supported and may be masked to 4:5; desktop Feed may show 1:1 with bars. TikTok’s official ads-spec HTML was not reachable when checked. This page does not treat 9:16 / 15s as a TikTok-official pair.

Twelve keys lotgen.ai pastes: Google, BytePlus ARK, fal, evolink, Kling, OpenAI, MiniMax, Black Forest Labs, Alibaba, ElevenLabs, xAI and Runway. Luma and Sora are models on fal or evolink, not keys. BytePlus ARK is the console that issues the Seedance key, and the name to put on the invoice.

What “best for ads” actually means here

“Best” on this page is not a cinematic quality score. This page has no takes to score. Best for ads, for an agency or a performance team, is the stack that produces an approved variant at the length and aspect the destination will run, at a dollar figure you can put on a media plan.

Five things move that number more than a beauty-contest ranking.

1. Cost per approved variant, not cost per generated second. List $/s is the sticker. The invoice is list × billed seconds × the takes you run before someone signs off. A cheap second that you throw away three times is not cheaper than a dearer second you keep.

2. Retries. A missed label, a bad hand, a talking-head that does not hold: you pay for the miss and the keep. Public provider pages checked for this page do not publish a keep rate. 3× and 5× below are a model you set.

3. Aspect ratios the destination actually uses. Meta Feed’s published recommendation is 4:5. 9:16 is supported and may be masked. Desktop Feed may show 1:1. A generator that documents 9:16 and 1:1 lets you quote a second render of the keep. A generator that only lists 16:9 and 9:16 (Veo 3.1) does not document 1:1. Crop-then-check is a different dollar line from a full re-generation.

4. Native audio. In-feed sound is optional on Meta’s Feed spec and recommended. A model that always bills audio (Veo 3.1 Gemini: “always on”) is a different meter from a model with a silent row and a native-audio row (Kling 3.0: 0.6 vs 0.9 Units/s at 720p). Do not compare a silent Kling cell to an always-on Veo cell and call it a fair ad-stack compare.

5. Bulk testing. Fifty prompts, one aspect, a review round: the interesting number is the batch total at your attempt slider, not the prettiest demo clip. The 50-variant section is that model. It is not a lotgen run.

This page does not repeat vendor “AI vs human UGC creator” savings claims as fact. If a vendor page advertises that gap, it is a vendor claim. Cite it only as such, with the human invoice in hand.

Cost per approved 15s variant across the twelve keys

Formula, so the page still works after a lab changes a cell:

billed_seconds     = clamp(requested_seconds, provider min, provider max)
cost_per_take      = list_rate × billed_seconds
cost_per_approved  = cost_per_take × attempts     # one aspect, before review

For BytePlus ARK the official meter is tokens, not a printed $/s. The comparable list_rate is computed from BytePlus’s published formula, then the same multiply.

15-second take, audio on where the key publishes a split. Same planning length as the short-answer table.

Closest official length when 15 seconds is not one take

Worked Kling 3.0 API, 720p, native audio, 15s

0.9 Units/s × $0.14 = $0.126/s
$0.126 × 15s = $1.89 per take
$1.89 × 3 attempts = $5.67 per approved variant

Worked Veo 3.1 Fast 720p, honest 8s cell (Gemini, audio default)

$0.10/s × 8s = $0.80 per take
$0.80 × 3 = $2.40 per approved 8s clip

That $2.40 is not a 15-second variant. It is an 8-second clip. A stitched two-clip model is in the 50-variant section, labeled as a model.

Worked Alibaba Wan 2.7 t2v International 720P, 15s

$0.10/s × 15s = $1.50 per take
$1.50 × 3 = $4.50 per approved variant

Official duration is 2 to 15 seconds. This cell is a billed length the API accepts, not a stitch.

BytePlus ARK Seedance 2.0 Fast 720p, list 5.6, 15s output, no input video

Official formula: (input seconds + output seconds) × width × height × frame rate / 1024. 24 fps, 1280×720, plus the one closing frame ARK bills (the product measured it on five real jobs; the page’s own examples reproduce without it, so this cell sits a cent above the page):

tokens = (15 × 24 + 1) × 1280 × 720 / 1024 = 324,900
324,900 × 5.6 / 1,000,000 = $1.82
× 3 attempts = $5.46

The same ModelArk pricing page still prints a time-limited 25% off on 480p/720p Seedance 2.0 Fast through 2026-10-07 14:00 UTC+8 (about $0.09/s at 720p). This table uses the list 5.6 rate. That window is dated copy, not a deadline on this page.

MiniMax dollars. PAYG list on the day checked: Hailuo 2.3 $0.28 / 768p 6s, $0.56 / 768p 10s, $0.49 / 1080p 6s. The video-package page still bills Hailuo 2.3 in points (768p/6s = 1 pt, 768p/10s = 2 pt, 1080p/6s = 2 pt). A 5% Standard package ($1,000 / 3,760 points) infers about $0.28/point. That inference matches PAYG 768p 6s and 768p 10s. It does not match PAYG 1080p 6s ($0.49). Use PAYG as list.

1080p, audio on, 15s where the official page prints 1080p

evolink’s Kling 3.0 page still tells readers official Kling is $0.089/s. Kling’s own page on the day checked is $0.084/s silent / $0.126/s audio at 720p. evolink’s “official” figure is wrong. Use it only as proof that published numbers disagree.

Sibling compare (live): Kling vs Veo cost owns the 8-second Kling vs Veo dollar walkthrough. Sibling Kling credits (live): Kling AI pricing explained. The cost pillar: how much AI video generation costs. The UGC cost page: AI UGC ads cost. Veo pricing: Veo 3 cost per second.

Capabilities that matter for ads

Only claims with a link to that provider’s docs. If a cell is “not printed,” this page does not fill it from a reseller, a remembered number, or a lotgen run. The one exception is labeled: where Runway prints nothing, lotgen’s own test of the route is given as measured by lotgen.

Duration. A 15-second Reels-shaped brief is a different job from an 8-second bumper.

Audio.

Aspect (9:16 and 1:1 from provider docs). Meta Feed’s 4:5 recommendation is in the short-answer note. This table is the generator side.

Text-on-screen. None of the official pages checked for this page publish a text-rendering SLA, a “legible end-card” guarantee, or a claims-safe on-pack mode. Alibaba’s t2v API has an optional watermark flag. That is a logo overlay, not a promise that product copy will be pixel-true. If the brief cannot be wrong (legal claims, on-pack ingredients), a keep-rate model is the wrong QA bar. This page does not invent a winner.

Talking-head vs product. This page has no takes. Do not rank Kling, Veo, Wan, or FLUX on faces, hands, or product-in-hand. Same-prompt footage is a compare still to come, and it needs a real run first. Avatar vendors that sell a generated actor are the aside in the previous section, not a row in these tables.

Product fact (not lab marketing): US lotgen workspaces cannot use BytePlus ARK direct. The Seedance row still belongs for keys that can paste it. US readers take fal/evolink Seedance, at marketplace rates.

A 50-variant test: total cost by provider and attempt rate

This is a worked model, not a lotgen.ai run. No prompts were submitted. No takes were kept. The arithmetic is the formula in the cost section, times 50 approved variants, one aspect.

Assumptions you can change: 50 approved variants, one prompt family each, 15-second billed length where the API will take 15, audio on / default, one aspect, attempts in {1, 3, 5}. Quality is not in the table.

Rounded to the nearest dollar in the 50-variant columns so a media plan can scan them. Exact cents used: Kling 720p audio 3× = $283.50; ARK 2.0 Fast 3× = $272.92; BFL hd 3× = $382.50; Alibaba 1080P 3× = $337.50; MiniMax 1080p 6s 3× = $73.50. Recompute from the one-take cells if you need invoice-level cents.

Worked Kling 3.0 API, 720p, native audio, 15s, 50 approved, 3 attempts

$1.89 × 3 attempts = $5.67 per approved variant
$5.67 × 50 = $283.50

Worked Veo 3.1 Fast 720p, 50 approved, 3 attempts, honest 8s cell (Gemini, audio default)

$0.80 × 3 = $2.40 per approved 8s clip
$2.40 × 50 = $120

That $120 is not a 15-second variant budget. It is fifty approved 8-second clips.

Stitched ~15s on Veo (model only). Two 8-second takes per attempt, then you cut. Not a lotgen run. Not a claim that the join will look like one shot.

Fast 720p:  $0.80 × 2 clips × 3 attempts × 50 = $240
Standard:   $3.20 × 2 clips × 3 attempts × 50 = $960

Edit time is extra and is not in these dollars. If the brief is “15-second Reels,” an 8-second-max model is a workflow tax, not a 15 × $/s line.

Worked Alibaba Wan 2.7 t2v International 720P, 50 approved, 3 attempts, 15s

$1.50 × 3 = $4.50 per approved variant
$4.50 × 50 = $225

Sora 2 is a model on fal/evolink, not the OpenAI key. OpenAI’s own pricing page on 2026-09-18: Sora 2 $0.10/s at 720p (720×1280 and 1280×720), Sora 2 Pro $0.30/s at 720p, Pro 1080p $0.70/s. Batch is half. A 15s Sora 2 720p take is $1.50 if that marketplace row accepts 15 seconds. It does not belong on the OpenAI key.

Retries are not the only multiplier. Aspect re-renders and a client round still show up on an agency job. A common agency shape, labeled as a model: 40% of the 50 approved variants come back for 2 more takes, and the keep also needs a second aspect.

Kling 3.0 API 1080p native audio, 15s (model):

Same stack at Veo 3.1 Fast 720p, honest 8s (no stitch), one aspect, no client round: $120 for 50, $2.40 per approved 8s clip. Add a stitch and it becomes $240 before aspects and review.

Same stack at Alibaba Wan 2.7 t2v International 720P, 15s, one aspect, no client round: $225 for 50 at 3 attempts, $4.50 per approved variant. Add a second aspect of the keep: 50 × $1.50 = $75, total $300 before a client round.

BFL’s published table is useful here in a different way: draft t2v at hd is $0.06/s ($0.90 per 15s take) against $0.17/s full hd. Exploring variants at draft and paying full render on the keep is a documented BFL pattern, not a lotgen observation. Fifty draft takes + 50 full keeps at 15s hd: 50 × $0.90 + 50 × $2.55 = $172.50 before retries on the keep.

When Kling vs Veo vs others for ads

The live dollar walkthrough for an 8-second clip is Kling vs Veo cost. Kling credits vs API Units live on Kling AI pricing explained. This section does not redo those pages. It does not invent same-prompt clips; that compare is still to come.

Pick from official meters, not from a beauty contest.

The brief is ~15 seconds, one take. Official 15-second billed lengths on the day checked: Kling 3.0 (evolink 3-15s + lotgen’s own test), fal Kling v3 schema 3-15s, Alibaba Wan 2.7 t2v (2 to 15s), xAI grok-imagine-video (up to 15s), BFL FLUX 3 (up to 20s). Veo 3.1 is 4, 6, 8. MiniMax Hailuo 2.3 at 1080p is 6s only. Runway gen4.5 tops out at 10s on lotgen’s own test. Stitching two Veo 8s clips is an edit, not a list price.

The brief is 8 seconds, or you can live with 8. Veo 3.1 Fast 720p at $0.80 a take ($2.40 at 3 attempts) undercuts Kling 3.0 720p native audio at 8s ($1.008 a take, $3.02 at 3 attempts) on the live compare. Veo Lite 720p is $0.40 / $1.20. Veo Standard is $3.20 / $9.60. “Veo 3.1” without Fast / Standard / Lite is a 4× invoice gap at 8 seconds. Get the tier in writing. Those Veo dollars are Gemini audio default.

You need 9:16 from the generator docs. Veo 3.1, Alibaba Wan 2.7, BFL FLUX 3, xAI, and fal Kling v3 print 9:16 on the day checked; Runway gen4.5 takes 720:1280 on lotgen’s own test. Veo 3.1 does not list 1:1. Wan 2.7, FLUX 3, xAI, and fal Kling v3 do. Kling’s own public t2v HTML did not print video aspect. MiniMax Hailuo 2.3 t2v did not print 9:16 / 1:1. Meta Feed may still display 1:1 with bars even if you generated 9:16.

You need native audio on the clip. Veo 3.1: always on, no silent SKU on Gemini. Kling 3.0: optional, +50% at 720p / 1080p (0.6 to 0.9 Units/s at 720p). fal Kling: generate_audio. Alibaba Wan 2.7: BGM / SFX without an audio file, or pass audio_url. BFL: synchronized audio. Do not mix a silent Kling cell into an always-on Veo compare.

You are iterating cheaply before a keep. BFL draft t2v hd is $0.06/s against $0.17/s full. Veo Lite 720p is $0.05/s at 8s. MiniMax Hailuo 2.3 768p 10s is $0.56 PAYG. Those are official iteration rows. They are not a quality ranking.

You have a named-model client spec. “Use Veo” without Fast vs Standard, or “use Kling” without API Units vs website credits, is how invoices surprise people. Kling AI pricing explained owns the credit/Unit split. Kling vs Veo cost owns the 8-second Kling vs Veo dollar table.

You need a generated actor with a consistent face. That is the avatar category in the aside above. Not a twelve-key row.

Live calculator, with the agency / performance preset: 15 seconds, 1080p, audio on, 50 shots, 3 attempts, the client toggle on. Pick the tile that fits, Agency or studio or Brand or in-house team. The 3-attempt slider stays a model you set. Veo rows will clamp to 8 billed seconds.

How to quote before you spend

lotgen.ai is a control plane, not the point of this page. One place to generate takes across the twelve keys, see a real-dollar quote before spending, compare outputs side by side, share a review link, and drive it via MCP. Invite-only beta. You still pay the provider. lotgen does not resell Kling credits or Veo tokens.

Public seat copy on lotgen.ai/pricing on 2026-09-18: $99 Solo, Team $90 a seat from 3, $79 from 6. Budget seats and seconds as two lines.

A quote that holds up in a monthly recap is dollars, seconds, model, tier, attempts, on the keys you will actually paste.

1. Brief the variant: billed length the destination will run, aspect (4:5 Feed vs 9:16), audio on or off, model and tier.
2. Quote the take in dollars on those keys, before the job runs.
3. Generate across more than one key when the brief allows it. Same prompt, two invoices, pick a keep. This page has no side-by-side render to show you. The same-prompt compare will, when a lotgen run exists.
4. Compare the takes with the quote still attached.
5. Share a review link so a client or brand manager signs off without a credit screenshot.
6. Approve, then spend the aspect re-render and the paid media.

Request an invitationSeats and pricingA dollar quote on your own keys before a 50-variant board runs.

Brand / in-house teams can use the same form: request an invitation. Seat and plan detail: /pricing.

This page is not a live lotgen.ai invoice and not a 50-variant test. Replace the model with real quotes before you treat a cell as a forecast.

Prices checked on 2026-09-18: how to re-check yourself

Refresh this page on each major model release, minimum quarterly.

1. Kling Units and per-second rows: kling.ai/dev/pricing. Confirm 1 Unit = $0.14 for video, then the Kling 3.0 720p/1080p/4K cells with and without native audio. Re-read duration and aspect on Kling’s own video API docs; this page used evolink’s 3-15s copy, fal’s 3-15s schema, and lotgen’s own test.
2. Veo 3.1 Gemini per-second rows: Gemini API pricing, section “Veo 3.1”. Confirm Standard / Fast / Lite. Google’s page was last modified 2026-09-16 when checked. These tables use Gemini audio default, not Vertex silent.
3. Veo 3.1 lengths and aspect: Veo 3.1 generate (4, 6, 8 seconds; 16:9 and 9:16) and Gemini Veo guide (audio always on; no 1:1 listed).
4. BytePlus ARK: ModelArk pricing. Confirm Seedance 2.0 Fast 5.6 USD/M tokens (list), whether the 25% off window is still printed, and 1080p support. Recompute tokens with the published formula. Do not paste a remembered 4.2 promo rate as list.
5. fal Kling: fal Kling Video v3 Standard. Marketplace. Confirm $0.126 audio on, duration 3-15, aspect 16:9 / 9:16 / 1:1. Do not paste fal Fast Veo over Google Fast.
6. evolink Kling: evolink.ai/kling-3. Marketplace. Ignore their “official Kling $0.089” line unless Kling’s own page has moved to that number.
7. MiniMax PAYG dollars: pricing-paygo. Use PAYG as list. Points stay on pricing-video. Hailuo 2.3 still not 15s in one take.
8. Black Forest Labs: docs.bfl.ml/quick_start/pricing and FLUX 3 overview for the 20-second cap and aspect list.
9. Runway: Dev pricing. Confirm gen4.5 at 12 credits/s × $0.01 and whether a duration cap is now printed on that row. Re-read the gen4.5 duration and ratio enums on the Dev API; until they are printed, the 10s cap and the two 720p ratios here are lotgen’s own test.
10. xAI: grok-imagine-video, 1.5, video capabilities (up to 15s), generation (1:1 / 16:9 / 9:16). Do not treat 1.5 audio input (preset voices) as a native soundtrack SKU.
11. OpenAI Sora (model on fal/evolink): OpenAI API pricing. Sora 2 $0.10/s at 720p on the day checked.
12. Alibaba: Model Studio pricing and text-to-video API. International wan2.7-t2v is $0.10/s 720P and $0.15/s 1080P. Duration 2 to 15s. Ratios include 9:16 and 1:1. Do not paste evolink Wan as Alibaba official.
13. Meta ad specs: Ads Guide, Facebook Feed video. TikTok’s own ads-spec page was not reachable when this was checked; read it there before you plan a Reels cell. Open Ads Guide, Reels, if you need a Reels-specific 9:16 cell.

Recompute: $/s × billed seconds × attempts × approved variants. If the billed seconds are not a length that provider accepts, stop and check the model id.

Related on this site: Kling vs Veo cost · Kling pricing explained · how much AI video generation costs · AI UGC ads cost · Veo 3 cost per second · AI video cost calculator · how a render is quoted before it runs · seats and pricing · request an invitation

Request an invitationSeats and pricingA dollar quote on your own keys before a 50-variant board runs.

FAQ