best 16 ai video models

this is the clearest result in either of our model categories. gemini omni flash is first on artificial analysis for text-to-video and image-to-video, with and without audio, and first on lmarena across half a million votes. two independent vote pools agreeing this completely is unusual.

and it costs ten cents a second. google's own veo 3.1 standard is forty cents a second and now ranks tenth on the same board. google has undercut itself by 4× and beaten itself on quality at the same time, which tells you something about how fast this category is moving.

audio is generated natively in the same pass — dialogue, effects and ambience, synchronised. it could not top the with-audio leaderboards otherwise.

the limits are real and worth planning around: ten seconds maximum, 720p maximum, and it is still a preview. google also documents that character consistency degrades across cuts and pans, and that video references up to three seconds are accepted by the api but not correctly processed yet. for short-form social, ads and b-roll that is fine. for anything longer you are stitching.

seedance 2.0 is the strongest model on this list for multi-shot storytelling: physics that hold up, director-level camera control, and shots that feel composed rather than generated. it sits second on both artificial analysis and lmarena, behind only gemini omni flash.

it does fifteen seconds against omni flash's ten, and up to 1080p against omni flash's 720p, so for anything that needs length or resolution it is the obvious step up.

the price is the problem. $0.30 a second at 720p is triple the leader; $0.68 at 1080p is the most expensive tier of any mainstream model. and a detail worth knowing: you are charged the same whether audio generation is on or off, so switching it off saves you nothing at all.

then there is the legal position. seedance's launch drew a cease-and-desist from disney, an infringement allegation from paramount skydance covering star trek, south park and dora, a statement from the mpa's chief executive accusing bytedance of unauthorised use of us copyrighted works on a massive scale, and a senate letter demanding shutdown. bytedance paused the international rollout. none of that makes the model worse, but it is a real consideration for anything client-facing.

alibaba's tongyi lab shipped wan 2.7 in april with four modes: text-to-video, image-to-video, reference-to-video, and instruction-based video editing. that last one is rarer than it should be and genuinely useful.

it ranks third on artificial analysis for text-to-video with audio — above kling, above veo 3.1 — at a flat ten cents a second regardless of resolution. seedance sits two places higher and costs three to seven times more. if you are producing volume, this is where the maths lands.

it does 2–15 seconds, 720p and 1080p, all the standard aspect ratios, nine-grid multi-image input, first-and-last-frame control and character reference. native audio can either preserve the original or regenerate it.

one correction worth making loudly, because a lot of coverage gets it wrong: wan 2.7 is not open weights. alibaba's public releases stop at wan 2.2, which is apache 2.0. the sites describing 2.7 as a downloadable apache model are wrong, and if you need weights you are getting 2.2, not 2.7.

happy horse comes from alibaba's taotian group rather than the tongyi lab behind wan, which is why it shows up on leaderboards as a separate vendor. it launched in june and immediately placed fourth on artificial analysis for text-to-video with audio.

for context on what that means: kling's 4K tier costs $0.42 a second and ranks below it. happy horse does 1080p with native audio for $0.18.

the standout feature is multilingual lip sync — mouth movement matched to speech across languages. for dubbed content, localised ads, or any talking-head format that has to ship in more than one market, that is the specific thing most models do badly.

the catch is entirely commercial rather than technical: minimal brand recognition, thin western distribution outside aggregators, and no 4K. if you can get past not recognising the name, the price-performance here is close to unbeatable.

kuaishou's kling 3.0 landed in february and its differentiator is control. you can storyboard a sequence and specify per shot: duration, shot size, perspective, camera movement. element consistency via reference uploads holds subjects across those shots. nothing else here gives you that much direction.

it is also the only model on this list producing native 4K — not an upscale — at $0.42 a second.

native audio covers english, chinese, japanese, korean and spanish including regional accents, though at the 4K tier audio narrows to chinese and english with everything else auto-translated to english. that is an odd limitation to hit late in a project.

the honest problem is value. it sits sixth and ninth on the artificial analysis text-to-video board while costing more than models above it. you are paying for the control surface and the 4K, not for output quality. one more thing: the widely-repeated claim that kling 3.0 runs at 60fps does not appear in kuaishou's own announcement — the 4K endpoint is confirmed, the frame rate is not.

veo 3.1's audio is genuinely the best engineered on this list: 48khz, with dialogue lip-synced properly, effects and ambience. if the output is going anywhere near a professional audio chain, that number matters.

video extension is its other real advantage. native generations are four, six or eight seconds, but chaining up to twenty extensions gets you to roughly 148 seconds of continuous video — far beyond anything else here, all of which cap out between ten and twenty seconds.

it does 4K, comes with google's enterprise trust and indemnity story, and integrates with flow.

the awkward part is that gemini omni flash — also google's — ranks first on the board where veo 3.1 now ranks tenth, at $0.10 a second against $0.40. if you are on veo 3.1 standard today and don't need the length or the 4K, you are paying four times over for a lower-ranked model. the fast tier at $0.10 is the more defensible choice.

on the numbers this is one of the best-value models in the category: third on artificial analysis for image-to-video both with and without audio, beating models that cost three to four times more, with native synchronised audio including lip-synced dialogue.

buy it direct from xai rather than through fal — xai charges $0.08 a second flat, while fal charges $0.14 at 720p. that is a 75% markup for the same model.

the ceiling is 720p, with no 1080p option at all, and up to fifteen seconds.

the part that belongs in any honest review: xai faces multiple lawsuits over non-consensual sexualised deepfakes, reporting that the video tool produced sexualised imagery of a named public figure without being explicitly asked, a multistate attorney-general letter, and an amsterdam court threatening €100,000-a-day fines. safeguards added in january were reported to apply only to non-paying users. whatever you conclude about the model, putting it in a consumer-facing product is a decision to take with your legal team rather than your engineering one.

pixverse v6 ranks fourth on artificial analysis for image-to-video without audio, effectively tied with grok 1.5, at $0.09 a second for 1080p. for silent footage that is the best quality-per-dollar on this list.

the pricing model is also more honest than most: audio is a per-second uplift you can decline. seedance charges you for audio whether you use it or not; here, switching it off actually reduces the bill.

it ships more than twenty cinema camera controls and a multi-shot engine, and handles one to fifteen seconds at up to 1080p.

the tell is in the leaderboard split: its with-audio ranking is markedly weaker than its without-audio ranking, which points at the audio component being the weak part. treat it as a strong silent-video model that happens to offer sound, rather than an audio-video model.

ltx-2.3 is a 22-billion-parameter audio-video model from lightricks, and it is comfortably the strongest open-weight option here: leader among open weights on artificial analysis' image-to-video with audio board, twenty-second clips — the longest of any mainstream model — up to 4K, and 24 or 48fps.

it generates synchronised video and audio in a single forward pass, which most open models still can't do at all.

hosted on fal it is also cheap: $0.06 a second at 1080p, $0.24 at 4K, with a fast tier at two-thirds of that.

now the licence, because this is the single most misreported fact in the category — including on fal's own model pages, which state the model is apache 2.0. it is not. the actual licence file is the ltx-2 community license agreement, and it requires organisations with at least $10 million in annual revenue to obtain a paid commercial licence. it also prohibits use in products competing with lightricks' own, and specifies liquidated damages of double the licence fee for unauthorised commercial use. read the licence file, not the blog post.

runway's gen-4.5 generates natively at 720p, which in a field that reached 4K is a real limitation, and it appears in neither leaderboard's top ten. as a text-to-video model it has been passed.

what runway still has is everything around the model: aleph for video-to-video transformation, act-two for performance capture, a genuine editing timeline, and 4K upscaling. those are workflow tools that the model labs mostly haven't built, and for teams doing post rather than generation they remain the reason to be here.

runway has also become substantially a reseller. its pricing page now sells seedance 2.0, veo 3, veo 3.1, gemini omni flash and happy horse alongside its own model. that is a sensible read of where things are going, but it does mean the differentiator is the interface rather than the output.

one urgent operational note: gen-3 alpha turbo and gen-4 aleph are being sunset on 30 july 2026. if anything you run depends on those endpoints, that is days away, not months.

first, a correction: dream machine, ray2, ray3 and ray3.14 are all superseded. ray3.2 arrived in june and anything recommending luma dream machine is out of date.

luma's real differentiator is the colour pipeline. native 16-bit HDR and EXR export in ACES2065-1 is something no other model on this list offers, and for a vfx shop that has to composite and grade generated footage alongside real plates, that is the difference between usable and not.

it also offers up to sixteen keyframes per clip, which is the finest-grained motion direction available anywhere here, and modify video v2 handles up to twenty seconds at 1080p.

the gap is audio. luma generates none, integrating third-party tools instead, in a year when single-pass native audio became standard. pricing is also opaque — luma doesn't publish a per-credit dollar value, so the roughly $1.20 per five-second generation figure is an approximation rather than a quote.

hailuo has a strong reputation for motion quality and a large consumer following, and the model itself is competitive.

the pricing is the problem and it is worth being blunt about it. minimax sells 'video points' in packages starting at $1,000 for 3,760 points, rising to $6,000 — and the points expire after a month. the documentation gives deduction examples rather than rates, so working out what a project costs means reverse-engineering it from worked examples.

every per-second figure you will find for hailuo comes from third-party resellers, not from minimax. we are not going to publish those as if they were the vendor's prices.

there are consumer plans from $9.99 and a developer token plan from $10 a month, which are more approachable, but the api story for anyone building at scale involves those package minimums. capable model, hostile commercial model.

firefly video is not a competitive model on output quality and does not appear in any leaderboard we checked.

it is the only major video model with a clean commercial-safety story: trained on licensed and adobe stock content, with indemnification available. in a year when seedance drew cease-and-desists from disney and paramount and xai collected deepfake lawsuits, that is a product feature with real commercial value.

be precise about the audio claim, though. adobe offers audio translation, lip sync and sound-effect generation as separate credit-consuming features. that is not the same as single-pass native audio generation, and vendors are increasingly blurring the two.

the commercial structure is awkward for anyone small: consumer plans from $9.99 to $199.99 a month don't include api access at all, and the firefly services api starts around a thousand dollars a month. like the image model, firefly video increasingly routes to partner models rather than competing directly.

hunyuanvideo 1.5 fits in 8.3 billion parameters and generates a clip in about seventy-five seconds on a single rtx 4090. for a research team or a solo developer without a cluster, that accessibility is the point.

it is still widely used as a fine-tuning base, and the weights are genuinely downloadable.

but the base model is silent — audio comes from a separate hunyuanvideo-foley model — and it tops out at roughly five-second clips at 720p. against ltx-2.3's twenty seconds at 4K with single-pass audio, it is a generation behind for open-weight work, and tencent has shipped nothing new on this line in 2026.

the licence deserves a careful read and is the reason it sits this low. it is the tencent hunyuan community licence, not an osi-approved open-source licence: organisations above 100 million monthly active users must request separate permission at tencent's discretion, outputs may not be used to train competing models, attribution is required, and — the clause most people miss — the licence expressly does not apply in the european union, the united kingdom or south korea. for a european company that is a hard blocker, not a footnote.

this entry exists as a warning rather than a recommendation. openai is retiring sora in two stages: the consumer web and app experiences closed on 26 april 2026, and the api shuts down on 24 september 2026. developers on the videos api and sora 2 model aliases were formally notified in march.

the model itself is not the problem. sora 2 pro still sits fifth on lmarena, ahead of every veo variant, and its native audio was genuinely category-defining when it launched.

but a model with a published end-of-life two months out cannot be recommended for any new integration, and its low placement here reflects exactly that. capability is not the same as viability.

if you are already on it, treat migration as urgent. gemini omni flash is cheaper and ranks higher; seedance 2.0 and wan 2.7 are the closest matches for longer or higher-resolution work.

pika 2.5 handles physics interactions and synchronised audio, and its effects-driven approach is genuinely fun for social formats. that was enough to matter in 2024.

it is now absent from every leaderboard we consulted, and there is no meaningful public developer api — it is a subscription consumer app.

the free tier gives eighty credits a month and paid plans start at $8 with commercial rights, which makes it cheap to play with.

we would not build anything on it. a widely-referenced 'pika 3.0' appears to be projected rather than shipped, so we're not ranking on it.