The Best AI Video Model for Ads in 2026: Five Boards, Four Winners

I checked five public rankings of AI video models on 10 September 2026 and got four different number-one models. Artificial Analysis has Alibaba's Wan 3.0 first at 1240 Elo with Google's Veo 3.1 fifteenth, arena.ai has Gemini Omni Flash first on 668,045 votes, and llm-stats has Kling v3 first on just 1,394 votes. Pixazo, which sells one API onto all of these models, scores Seedance 2.0 first and marks Veo down at 934, and FilmBench, built with Beijing Film Academy faculty, also puts Seedance 2.0 first at 88.93.

Which model wins depends on the job, and leaderboard position barely enters into it. Run MiniMax H3 Max for volume testing at roughly $8 of compute per finished thirty-second spot, Kling 3.0 for multi-shot sequences, Seedance 2.5 for heavy product reference binding, and Gemini Omni Flash as the general default, which is Google's own documentation talking. That is more than the it-depends pieces gave you, with one caveat: MiniMax is a named defendant in US copyright litigation that survived a motion to dismiss, so read the docket first.

Which AI video model is best for ads in 2026?

Three things move that table, and none of them is a benchmark score. Start with whether the shot contains a real human face you did not generate yourself, because ByteDance's face policy blocks that route outright while Google's and Kuaishou's terms restrict it in different ways. Legible text in frame is the next gate, since no model in this set reproduces small legal copy or tight kerning reliably and no leaderboard even measures the attempt. Last and largest is your reroll budget, which swings on shot type far more than on which model you picked.

Why do AI video leaderboards disagree with each other?

All five were live on 10 September 2026. Read the sample column before the score column.

Two of the five are too thin to disagree with anything. llm-stats runs about 116 votes per model, which under a conservative mu-minus-three-sigma rating rewards whichever model collected the most votes, and Pixazo is a board run by a company selling API access to the models it ranks, on 450 matches a track and an undescribed rater pool. Weight by sample and the serious boards mostly agree: Gemini Omni Flash and Wan 3.0 near the top, Veo 3.1 well down both. The real problem is that the thing they agree on still does not predict what survives a thirty-second cut. FilmBench is the one worth reading, the only one of the five scored against film-school criteria by people who make films.

Artificial Analysis does not agree with itself across tracks. Its image-to-video board puts MiniMax H3 Max first at 1200 and Seedance 2.0 second while its text-to-video board puts Wan 3.0 first and MiniMax third, one rating system landing on two winners because the task changed. The scales do not line up either: Elo boards here run roughly 1000 to 1500, while llm-stats runs 1750 to 1950 on a scaled TrueSkill rating whose scaling constant it never publishes, so comparing Kling's 1934 against Wan's 1240 compares nothing.

arena.ai publishes error bars, and its top six sit at 1515±15, 1511±10, 1494±19, 1494±17, 1491±19 and 1482±12, close enough to make the ordering arbitrary. Artificial Analysis publishes no confidence intervals and no vote counts, so when first through fifth run 1240, 1239, 1235, 1228 and 1222, nothing says whether 18 points means anything. Dehghani and seven co-authors called this the benchmark lottery in 2021, showing across NLP, vision, retrieval and reinforcement learning that "the relative performance of algorithms may be altered significantly simply by choosing different benchmark tasks". Nobody I can find has applied it to video, where it bites hardest.

Pages ranking for this query quote ranks that were never true. On 10 September 2026 Replicate's collection page said "Runway Gen-4.5 is the top-rated video generation model, ranked #1 on the Artificial Analysis text-to-video benchmark", while Wan 3.0 held first that day and Gen-4.5 was nowhere near the top group. Others say Seedance 2.5 leads that board, and 2.5 is not listed on it in any form.

Does a higher leaderboard rank mean better ad creative?

No. A high rank tells you a model wins short close-up clips, which is a different question from the one an ad brief asks.

The boards are made mostly of faces

More than 60% of arena test cases are portrait or talking-head prompts, per Wireflow's analysis, the only sourced number I have on arena's prompt mix. A model tuned for close-up faces wins a board built out of close-up faces, which tells you little about a product rotating under a key light. I have not found a source establishing that any specific model is weaker on close-ups than its rank implies, so I will not assert it, though the composition point stands: the test set is not shaped like an ad brief. Voting is gated behind a short watch timer, practitioners say, a number of seconds I could not confirm; either way, judging a clip on its first few seconds rewards high contrast, shallow depth of field and a slow push-in. That register is what most of my de-slop pass exists to remove, and drift at second nine never gets sampled.

"The Leaderboard Illusion" documents the adjacent problem, labs testing many private variants and publishing only the best, which inflates their apparent position. It covers Chatbot Arena and text models, and since video arenas have no equivalent of LMArena's Style Control, carrying the argument across to video is my own inference and not the authors' claim.

No board measures on-screen text or multi-shot

Signage, prices, packaging type and legal lines are not optional in commercial creative, and none of these boards scores them. Seedance's on-screen text is unreliable for exactly those cases while subtitles and captions render better, and ByteDance has never claimed otherwise. Generate a clean motion plate and composite the brand assets in post, which is why my prompts end by reserving space for the price or logo overlay.

Two 2026 papers make the multi-shot problem concrete, and neither is cited in the articles ranking for this query. FilmBench, 27 July 2026, built its 1,169 prompts by reverse-engineering award-winning films and reports "a marked single- to multi-shot performance drop that widens for weaker models": 7.9 points on average, the strongest losing about 2.3 and the weakest as much as 22.8. MovieGrid, 6 September 2026, reports inter-shot consistency around 0.59 against intra-shot consistency around 0.91 for its own post-training method, best-in-class results with no industry measurement behind them. Inside a shot, character consistency is close to solved; the moment you cut, it comes apart.

How much does one 30-second AI video ad cost?

How many generations it takes to get one keeper

This number decides your invoice, and it varies by shot type more than by model. Poppify's breakdown is the best public version I found, though it is a vendor page with no published sample or method.

invideo sells an AI video product and has every incentive to talk waste down, so its own numbers are worth reading. Its FAQ says roughly 25% of generated clips make the final cut and that 10 to 40 prompts per usable clip is common for ad-grade work, while a documented four-ad run on the same site reports "108 images and 103 videos were generated to land 49 used video clips", called there "roughly a 50% video-clip utilization rate": two invideo figures that disagree about the denominator.

The ratios get worse at the broadcast end. PJ Accetturo has described the Kalshi NBA Finals spot as taking three days and about $2,000, as the primary creator working with Kalshi's team, per the reporting. The counts circulating for it, 300 to 400 generations for 15 usable clips, trace to secondary coverage with no primary source, so the 4% keeper rate is an anecdote. Skill variance swamps the model difference: the same brief costs far fewer generations for someone who has built the shot before, which I argue in my guide to prompting Seedance 2.5 without burning credits.

The formula worth memorising

Cost per finished second = (raw dollars per generated second ÷ keeper rate) ÷ edit retention rate.

Both denominators are fractions below one, so both inflate the number. A model at $0.10 a second with a 25% keeper rate and 70% edit retention costs $0.57 per finished second, nearly six times the sticker price, and every per-second comparison table I opened while writing this quotes the numerator alone.

Cost per finished 30-second ad, modelled

These are modelled numbers: ten usable five-second shots per ad at four generations per keeper, which is fifty seconds of keepers out of two hundred seconds generated, for a thirty-second cut and already bakes in 60% edit retention, so read the ordering as the finding and the absolutes as an estimate. Keeper rate stays constant across models because I have no per-model data and the number belongs to the shot. Each per-second figure is one provider's rate for one tier, retrieved 10 September 2026: Veo from Google's pricing page, Seedance at BytePlus ModelArk's official rate via Cellcog, the rest from the Artificial Analysis price-per-minute columns.

Kling is the only price I could not confirm on a first-party page; the flag stays because the article recommends the model. The Seedance reference row assumes a five-second reference clip against each five-second output, which by the billing formula below doubles the billed duration, and that is the workflow you buy Seedance to run.

Across comparable tiers the spread is roughly 2x to 3x, reaching 10x only when you set a budget model against a flagship tier, so pick the tier before the vendor. The "Veo costs ten times MiniMax" framing collapses on Google's own price list, where Veo 3.1 Lite lands within a rounding error of the cheap Chinese models. diyai.io models a mostly non-overlapping set at three attempts per accepted shot and 70% edit retention, landing at Veo 3.1 Fast $15.43, Kling 3.0 Standard $15.58, Luma Ray 3.2 $30.86, Runway Gen-4.5 $37.03 and Pika 2.5 $51.43, and it is no cross-check, because the one model both tables contain sits at opposite ends of the two lists. I am also pricing from Artificial Analysis after telling you not to buy on its ranks, which holds up only because its price column aggregates list prices and inherits the provider spread below.

Why the same model costs twice as much elsewhere

Seedance 2.5 at 720p runs from $0.2312 a second on BytePlus ModelArk and Replicate to $0.4730 on fal, per Cellcog's cross-checked table, with WaveSpeed Turbo at $0.20 setting the floor and the whole 720p provider set spanning about 2.4x. One finished ten-second clip is $2.31 officially and $4.73 at fal, for identical output.

The billing formula matters more than the sticker. Seedance bills on `width × height × (input_video_duration + output_duration) × 24 / 1024` tokens, per apiyi's API documentation, so a reference video's duration counts toward the bill even though it is an input: a ten-second reference plus a ten-second output bills as roughly twenty seconds, and reference-to-video often costs more than text-to-video on the same model. Input images are not billed, and some resellers discount counted input seconds, fal at 0.6x, so check the rate card before you budget the workflow that made you buy the model.

The line nobody prices is editing labour

One itemised breakdown of a finished sixty-second AI spot, from AI Video Bootcamp, comes to $229.80 fully loaded. Editing labour at two hours and $50 an hour is $100 of it, 43.5% and the largest line, while video compute, priced on Veo 3.1 with audio and a 3x discard built in, is $72.00 or 31.3%, so the compute share shrinks on any cheaper model. AI Video Bootcamp sells a training course, so treat it as a publisher's unaudited itemisation.

Writing and directing appear at zero in that stack, as in every vendor comparison I have read, though they are the entire input. Accetturo's production costs cap around $2,000 a spot while his agency charges five figures, and cheap compute does not close that gap. Nielsen's 2017 study of nearly 500 campaigns put creative at 47% of sales lift against targeting's 9%.

AI video model comparison 2026: can it carry an ad?

Columns come from vendor documentation, official published rates where they exist and Artificial Analysis converted to seconds where they do not, and ranks from Artificial Analysis text-to-video with audio. All read 10 September 2026, with cells I could not confirm on a vendor surface marked n/a.

Runway, Luma and Pika sit outside the workflow I would build for an ad, because a brief needs reference control and native audio and each is weaker on at least one. Pika's last release is 2.5 from 2025; I do not know what that says about the company's plans, though a model with no 2026 successor is one to ask direct questions about before you build a pipeline on it.

Seedance 2.5, and the 4K claim that belongs to a different model

Is Seedance 2.5 native 4K? Not on any developer surface I could verify as of 10 September 2026. Hosts expose 480p and 720p consistently, 1080p arrived around mid-August with providers still disagreeing, and nobody I could check exposes 4K through an API. At the 23 June 2026 announcement ByteDance shipped 2.5 and separately upgraded Seedance 2.0 to native 4K; the English-language tech press ran the two together, while Chinese coverage kept them apart explicitly. MindStudio's explainer still says 2.5 "adds 4K output resolution".

ByteDance is not blameless. Its own consumer page is titled "Official Seedance 2.5: 4K & 30s AI Video Generator with Audio", while fal's page says 2.5 is 480p and 720p and 2.0 is the one doing 1080p and 4K, so the vendor's marketing surface and its API surface disagree and resellers have landed in different places. I cannot rule out a 4K route I could not reach, since ByteDance's ModelArk documentation needs a session to read. Confirm on the surface you are buying before promising a client a resolution.

What 2.5 is genuinely good at is reference binding: 50 references, 30 image, 10 video, 10 audio, against 2.0's nine, three and three. Give each reference one job and forbid everything else, in the shape `@Image1 controls only the exact espresso maker: preserve its matte cobalt-blue shell and copper button. Do not copy @Image1's studio background.` For a talking head the equivalent is `Her mouth moves only during her own lines`, which removes the most obvious tell in AI UGC. There is no `negative_prompt` field, so negatives go in as prose at the end, and reference numbering follows upload order, so a wrong upload order silently binds the wrong image to the wrong job. The rest is in reference binding and prompt grammar for Seedance 2.5.

At $0.2312 a second officially it costs more than every model I recommend except Veo 3.1, about 50% above Artificial Analysis's listed rate for Seedance 2.0 at 720p, though that sets an official rate against a reseller listing. Segmind measured wall-clock latency from 60.1 to 226.7 seconds, median 138.9, a 2.9x spread on identical requests because queue depth dominates, which rules out generating anything live in front of a client.

Gemini Omni Flash vs Veo 3.1, and the premium no leaderboard prices

Google's Gemini API video documentation says to "Use Gemini Omni Flash as your default model for video generation" and to use Veo 3.1 where "scene extension, last-frame control, or integration with legacy pipelines are required". Read that as a cost and latency recommendation from the vendor; it says nothing about quality.

What justifies Veo's price for brand work sits on no leaderboard. Google carries an output indemnity for Veo under its generative AI indemnified services terms, with a trademark carve-out that bites in advertising specifically, advertising being trade or commerce by definition, and no Chinese-hosted model here offers anything equivalent. If you are spending your own money on hooks, take the cheap model; if you are shipping a national spot for a client whose counsel reads vendor terms, the $60 you save per ad is not what is being priced. Artificial Analysis lists Veo 3.1 at $24.00 a minute against Omni Flash at $6.00, while Google's own list puts Veo 3.1 Lite at $0.05 a second at 720p.

Kling 3.0, the multi-shot option

Kuaishou's own comparison page claims up to 15 seconds, 4K "in supported workflows", and "multi-shot generation, custom shot timing, framing, angles and camera movement", with a Custom Multi-Shot mode where you "describe each shot separately and set its duration". Nothing else here markets that. You can approximate it in Seedance by tiling a prompt with timecoded segments and explicit cut markers, though no documented parameter backs that, so Kling is where you specify the cuts, and since FilmBench and MovieGrid both identify multi-shot as the hard problem, that matters for any ad longer than one continuous take. The page is a vendor ranking its own models and never mentions Kling 3.0 Pro sitting tenth on Artificial Analysis; I cite their product claims because I can check those, and ignore their ranking.

Wan 3.0 tops the biggest board and is not open source

Half the comparison pieces I read call Wan 3.0 open source, when as of 10 September 2026 I could find no published weights, no inference code and no ComfyUI node for it anywhere, and the Wan-AI organisation on Hugging Face still shows Wan 2.2 under Apache 2.0 as the last open flagship. I checked Hugging Face and not ModelScope, where Alibaba often publishes first, so check both, and check again before planning a self-hosted deployment, because this team has released weights before and may again. The licensing elsewhere is no cleaner: HappyHorse gets promoted as the number-one open-source model while fal, its official API partner, states it will be closed source and not licensable, and MiniMax H3's open weights need a separate application for the USA, EU, UK and South Korea.

MiniMax H3 and H3 Max, the cheapest credible option

At $2.40 a minute on Artificial Analysis, H3 Max is roughly a tenth of Veo 3.1 Standard and sits third on text-to-video and first on image-to-video, so volume testing is where I would spend. Motion's Creative Benchmarks 2026 shows hit rates rising only from 4.0% to 8.8% across spend tiers while weekly testing volume rises 6.7x. That is a correlation inside one vendor's customer base, where higher-spending accounts differ from lower-spending ones in every way that matters, so it cannot establish that volume causes hit rate. My argument that creative quality dominates points the other way, and I would rather leave the tension visible.

Price one more thing alongside the $0.04. MiniMax is a defendant in Disney Enterprises v. MiniMax, No. 2:25-cv-08768 in the Central District of California, where on 22 May 2026 the court denied its motions to dismiss on both personal jurisdiction and the merits, the most consequential pro-plaintiff ruling yet against a video model maker. No US court has ruled on training and fair use for video, so this is exposure of unknown size, and if your client's legal team screens vendors on litigation posture, this is the one that fails.

The Sora API shutdown date is 24 September 2026

OpenAI notified developers on 24 March 2026, closed the consumer app on 26 April, and listed the Videos API and sora-2 for removal on 24 September with no recommended replacement listed. Wikipedia's account states plainly that OpenAI gave no specific reason in the notice, reporters have attributed it to compute reallocation toward coding and enterprise products, and the download-decline figures circulating do not appear in either source, so I have left them out. OpenAI signed a three-year licensing deal with Disney in December 2025 covering more than 200 characters and shut the product down inside a year anyway, so a pipeline with one model hard-coded into it is exposed to a decision you get no visibility into.

Gemini Omni Flash is the closest general-purpose replacement, with Seedance 2.5 where Sora carried reference-bound product work and Kling 3.0 where the pipeline relied on multi-shot, and the job table at the top of this piece maps the rest. Keep the storyboard and script layer separate from the generation call, but prompt grammars are incompatible between these models, so budget a rewrite of the prompt layer with any swap.

Six constraints that decide the model before quality does

1. Who has to be told this is AI. AI Act Article 50 disclosure duties have bound EU audiences since 2 August 2026, with the Article 99 penalty tier at EUR 15M or 3% of worldwide turnover; the 50(2) provider-side marking grace period ends 2 December 2026, and the deployer duties binding an advertiser had none. Chinese-hosted models carry CAC labelling obligations effective 1 September 2025, a visible label plus metadata, which matters because three of my five recommendations are Chinese-hosted. A synthetic presenter has to be a demonstrator: FTC endorsement rules do not care that your customer is fictional, and inventing a purchase or a before-and-after is the fastest route to a Section 5 problem. Meta's ad standards and Google's misrepresentation policy carry account-level penalties, Google after at least seven days' warning.
2. Can you upload a real actor's face? With Seedance, no: ByteDance applies technical controls blocking video made from images or video containing real faces, as their own BytePlus blog describes, while Google's and Kuaishou's terms restrict the same route in different ways short of an outright block. The sanctioned routes on Seedance are real-person verification and likeness authorisation through the ModelArk console, or ByteDance's library of over 10,000 virtual human assets. Everyone shipping AI UGC generates a synthetic actor in an image model first and feeds that face to the video model, which fixes your pipeline shape before you compare a single output.
3. Will it render text you can legally ship? No model here reliably reproduces small legal copy, tight kerning or branded type.
4. How long will you wait, and how much does that vary? Segmind's 2.9x variation on identical Seedance requests is queue depth, outside your control, and while that figure is one provider on one model, anyone quoting a single generation time is quoting a median and calling it a spec.
5. What happens when moderation rejects a generation you already paid for? Segmind saw requests referencing specific film stocks or branded looks rejected after generation completed, a cost line no comparison prices.
6. Can you get the weights? Not for Wan 3.0 or HappyHorse, and MiniMax H3 needs a filed application, so the top of the most-cited leaderboard is closed to anyone who has to self-host.

Does using AI video hurt ad performance?

Exner, Hartmann, Netzer and Zhang's study of 4,633 sibling ads across 369 million impressions found AI and human creative at parity once you control for campaign, with the penalty attaching to images that read as AI rather than to images that are AI. Their looks-like-AI coefficient is -0.3759 uncontrolled and -0.3468 with full controls, in a working paper still awaiting peer review. Two of their cues are actionable: intense colour saturation signals AI generation to consumers, and medium-to-large faces read as human.

The demand side points the other way. An NYU Stern finding carried in version 2 of the IAB's AI Transparency and Disclosure Framework, published 18 August 2026, puts the click-through cost of an AI label at 31.5%. Gartner's survey of 1,539 US consumers, fielded October 2025 and released March 2026, found half preferring brands that avoid generative AI in consumer-facing content. And Coca-Cola's 2025 AI Christmas ad took the year's loudest public backlash while System1 scored it 5.9 stars, the top of their scale, a tension worth sitting with.

Model choice solves none of that, though the parity result being conditional does change what you optimise for, since it makes controlling how the work looks the lever you actually hold. I unpack the practical version in why ads that look AI-made underperform ads that are AI-made.

One model or several, and what I would run next week

Run two models across the funnel and keep them out of the same cut. The cheap one handles concept testing, where the footage is disposable, and the expensive one takes the winner, regenerated end to end. Today that pairing is MiniMax H3 Max plus whichever of Seedance 2.5, Kling 3.0 or Gemini Omni Flash matches the constraint binding your brief.

Do not intercut two vendors' output inside one finished spot. Inter-shot consistency is unsolved inside a single model, and across models no reference system carries at all, so skin, colour science, motion cadence and grain all fail to match and you pay it back in the grade. With editing labour already the largest line, that is an expensive way to save $30 of compute.

Pay for the expensive model when the shot has a product that must not deform, a face that must stay the same across cuts, or a client who will watch it frame by frame. Skip it on hooks, B-roll, establishing shots and anything answering a question about the concept before the craft matters, because running five variations of an unproven concept teaches you only which shade of a bad idea is least bad.

For a spot next week I would storyboard in stills and get sign-off before a single render exists, generating the boards in an image model and running a de-slop pass over the intense colour saturation the sibling-ad paper names as an AI cue. The rest of that pass, the plastic skin, the beauty-filter smoothing, the teal-orange grade, is my judgement and no finding of that paper. The stills step alone moves generations per keeper from roughly ten to roughly four in my own runs.

Draft at 480p, about 0.44 times the price of 720p on the official rate card, and approve at 720p. Change one variable per retry. Region and clip edits help once the rest of the frame is approved, though a Seedance edit bills against the reference video's duration, so several small edits can cost more than one clean reroll. Budget on the assumption that hands, lip-sync and walking shots cost six to ten generations each and static product beauty shots one or two, and put the cheap model on everything outside the hero shot. Then run the frozen-frame QA before anything ships, which I have written up as a six-point checklist for catching AI tells before a client does.