New AI video generators: 8 models to watch in 2026

Gemini Omni is Google's answer to a simple problem: generating a clip and then fixing it used to be two separate jobs. Omni does both in one conversation. You animate a photo, ask for a change, then another, and the model keeps the scene. Google calls it its latest video editing and generation model, and it will replace Veo in the Gemini app. The current version, Omni Flash 1.2, arrived in 2026 and makes clips of up to 15 seconds, up from 10 with Omni 1.1, with native audio.

You reach it two ways. In the Gemini app (see our Gemini review), video generation is listed from Google AI Plus, at $4.99 a month. In Google Flow, Google's video studio, everyone gets 50 free Flow credits a day, with or without a subscription. Omni Flash 1.2 costs 3 credits per second at 720p and 1 credit per second at 360p. So the free allowance buys one 10-second clip at 720p a day (30 credits), with 20 credits left for a 360p draft. Google AI Pro, at $19.99 a month, adds 1,000 Flow credits: about 33 ten-second clips at 720p.

What stands out in our comparison is the price of a draft. At 1 credit per second in 360p, you can try five versions of a 10-second idea for 50 credits, then pay for the 720p take you keep. Clips longer than 10 seconds are for Plus, Pro and Ultra subscribers, and each extra second costs double.

Who it's for: a marketer or a creator who wants to try a new Google model for free and refine it by talking to it. Skip it if you need 1080p straight from the model: Omni Flash 1.2 generates at 360p and 720p, and 1080p comes from an upscale reserved to paid plans.

The risk: Google retires models fast. Veo 3.1 Fast, Veo 3.1 Quality and Omni Flash 1.1 leave Flow in 2026, so a workflow built on them has to move. Without a subscription, Flow may also refuse video during peak hours, roughly 2 PM to 5 PM UTC. In the Gemini app, limits are compute-based and refresh every 5 hours up to a weekly cap, which makes your real number of clips hard to predict.

Kling is the video model from Kuaishou, and its two current versions are Kling 3.0, with synchronized audio and storyboarding, and Kling 4.0. What's new in 4.0 is length. Kling describes it as supporting native 30-second generation and multiple keyframes. Thirty seconds in one take means fewer joins between clips, and fewer chances for a face or a product to drift from one generation to the next.

The pricing is built around a first-month discount. Standard costs $6.99 the first month, then $8.80 a month on monthly billing, for 660 credits a month; on that plan, 100 credits cost $1.06. Pro costs $25.99 the first month, then $32.56 a month, for 3,000 credits. Billed yearly, Standard costs $79.20 for the year, about $6.60 a month for the same 660 monthly credits, the lowest monthly rate on Kling's paid plans. The free Basic plan lets you try the models, but commercial use and removal of the Kling watermark are listed for the paid plans.

Two rules work in your favor. Failed generations don't consume credits. And credit packs bought separately stay valid for 2 years, while the monthly credits of a plan don't roll over. If your volume swings from month to month, a small plan plus a pack wastes less than a big plan you half use.

Who it's for: a creator or a brand that needs longer continuous shots: a product reveal, a walk-through, a 15 to 30-second ad. Skip it if you need a clear cost per second before you subscribe: the consumer plans price credits, and the per-second rates Kling publishes are for its developer API ($0.084 per second for Kling-V3 in standard mode, without audio).

The risk: customer service. Kling's public reviews average 1.5 out of 5, and reviews written in 2026 describe credits that vanished and subscriptions that renewed without the user meaning to. Turn off auto-renewal the day you subscribe if you only want one month. Kling 4.0 also runs inside Runway and Luma, if you'd rather pay a vendor you already use.

Seedance is the video model family from ByteDance, the company behind TikTok and CapCut. Seedance 2.0 spread to the big hubs in 2026 (Runway, Luma and LTX Studio all list it). Seedance 2.5 followed, and Dreamina, the creative app on CapCut's site, sells it direct next to a lighter Seedance 2.0 Mini. Runway, Luma and Higgsfield carry Seedance 2.5 as well, which tells you how much demand there is for it.

Buying it at the source gives you the clearest math. Dreamina Basic lists a renewal price of $15 a month for 1,575 credits, and Seedance 2.5 costs 37 credits per second at 720p. That's about 42 seconds of Seedance 2.5 a month at 720p, before retakes. Dreamina's own pricing guide spells out the trap: 300 generated seconds use 11,100 credits with Seedance 2.5, against 1,500 with Seedance 2.0 Mini. In the US, an introductory offer has run the first month at $0.29 with the same 1,575 credits, paid by credit card, debit card or PayPal.

The source is not always the cheapest place, though. On Luma's Plus plan ($30 for 10,000 credits), Seedance 2.5 at 720p costs 78 credits a second, about $0.23, against about $0.35 a second at Dreamina's $15 renewal. Dreamina wins on small volumes and on the first month; Luma wins once you'd pay for a bigger allowance anyway.

In our comparison, the smart move is to split the work. Draft on Mini (5 credits per second at 720p), then render the take you keep on 2.5 (37 credits per second). The trial-and-error part then costs about seven times fewer credits.

Who it's for: a social video creator or an e-commerce brand that wants Seedance at the source, with cheap drafts. Skip it if you need predictable delivery times for a client deadline.

The risk: queues and filters. Creators in 2026 describe "brutal" queues during US working hours, and one says they lose about 40% of their credits to face blocks and content filters. Dreamina's public reviews average 1.5 out of 5, mostly about billing, including a one-time credit purchase that kept charging every month. Read the renewal line at checkout: the $0.29 month renews at $15. For CapCut itself, see our CapCut alternatives.

Runway is on this page for what it changed in 2026: it stopped being a one-model shop. Next to its own Gen-4.5 (text and image to video) and Aleph 2.0 (precise editing of existing footage with a prompt), with one-click Apps that relight a scene, swap a backdrop or remove an object, its plans now list Kling 4.0 and Seedance 2.5. One subscription, one credit balance, and a credit price per 5 seconds for each model.

The grid is easy to read. Free gives 125 one-time credits. Standard costs $15 a month, or $12 a month billed annually, for 625 credits a month, which Runway itself converts into 52 seconds of Gen-4.5. Gen-4.5 costs 60 credits per 5 seconds and Aleph 2.0 costs 140, so a Standard month buys four 5-second Aleph edits. Pro costs $35 a month, or $28 billed annually, for 2,250 credits.

What we like in our comparison: Runway shows the total cost before you generate, and it answers the rights question plainly. What you create is yours on every plan, the Free plan included.

Who it's for: a creative team that wants several new models, an editor for real footage and a single bill. Skip it if you'll only ever use one outside model: compare that model's own subscription before you pay Runway's credits for it.

The risk: credits burn fast. Among individual plans, unused credits roll over (for one month) only on Max, at $95 a month or $76 billed annually; the Team plan rolls them over too. Credits you buy on top of a plan never expire, from a minimum of 1,000. Runway's public reviews average 1.1 out of 5, and reviews written in 2026 complain that credits run out quicker than expected, with one month's allowance spent on a single short film. Start on Standard, note what a usable clip really costs you in credits, then decide. For the established ranking, see our best AI video generator page.

Grok Imagine is the image and video generator built into Grok, xAI's assistant. You write a prompt or upload a photo, get a clip, then restyle and edit it in the same thread. xAI's developer docs give clips of up to 15 seconds. Its interest is less the model than the bundle: video comes with an assistant you may already pay for, and the same subscription now carries X benefits too (fewer ads, a checkmark).

Pricing follows that bundle. Plus, the entry plan, costs $8 a month and lists Imagine in a usage pool shared across Grok's products. Premium costs $30 a month, with 4x the usage of Plus and "stunning AI images & videos" in its list. 1080p video, in 15-second clips, appears on Super, at $100 a month, with 15x the usage of Plus. The plans page states usage in multiples of Plus, so you only learn your real number of clips by spending it. Developers pay by the second instead: grok-imagine-video costs $0.05 a second at 480p and $0.07 at 720p on the API.

Our Grok review covers the assistant. For video, check what your tier lists (Premium names videos explicitly), then watch how fast the shared pool empties before you move up. If the limits put you off, our Grok alternatives compare the other assistants.

Who it's for: someone who already uses Grok or X and wants quick social clips without another subscription. Skip it if you need to quote a client budget: usage multiples don't convert into seconds of video.

The risk: rules that change without warning. Free access to Imagine was cut in 2026 with no announcement, and the complaint that stood out among users wasn't the price but the silence: "The issue isn't monetization. It's the lack of communication." In the same thread, a paying user reported about 10 video generations in a whole day.

Luma is the company behind Dream Machine, and Ray3.2 is its current video model, listed first on its pricing grid. What it brings is a finishing format: Ray3.2 renders in standard dynamic range, in HDR at 2x the base credit cost, or in HDR with EXR files at 3x, for teams that color-grade their footage. Like Runway, Luma opened its plans to outside models: the same grid lists Seedance 2.5, Seedance 2.0 and Kling 4.0 next to Ray.

Individual plans start with Plus at $30 a month, for 10,000 credits, with commercial use. Ray3.2 costs 100 credits per 5 seconds at 720p and 400 per 5 seconds at 1080p, so Plus buys 100 five-second clips at 720p, or 25 at 1080p in standard range. Pro costs $90 a month for 40,000 credits and Ultra $300 for 150,000. Billed yearly, Plus drops to $25 a month ($300 a year).

In our comparison, Luma makes sense once you know you'll deliver 1080p HDR: at that level, the credit math weighs more than the subscription price. For quick drafts, Ray3.2 at 360p costs 20 credits per 5 seconds, which stretches a Plus plan a long way.

Who it's for: a post-production team or a filmmaker who wants HDR and EXR straight from an AI model. Skip it if you want to try before paying: every plan on Luma's pricing page is paid, from Plus up.

The risk: speed and support. Luma's public reviews average 1.4 out of 5, and reviews written in 2026 complain that output quality drops once you subscribe, while credits keep burning. Pay one month before you commit to a year.

LTX is the video model family from Lightricks. LTX-2.5 came out in 2026 with downloadable weights, and it's also available on every plan of LTX Studio, Lightricks' web app with storyboards, camera controls and node-based automation called Flows. That combination is what makes it worth watching: a recent model you can run on your own machines, and a finished app when you don't want to.

Read the license before you say "open source". The code on GitHub is under Apache 2.0, but the models are released under their own license, the LTX-2.x Community License. Entities with annual revenues of at least $10,000,000 must obtain a paid license, and revenue is counted across affiliates. Below that line, you can use the weights for free. That's open weights under a community license, which is different from an open source license.

In LTX Studio, the free plan gives 800 one-time credits under a personal use license. Lite costs $12 a month billed yearly ($144 a year, or $15 month to month) for 8,000 credits a month and removes the watermark, but it carries the personal license of the free plan. The commercial use license comes with Standard, at $28 a month billed yearly ($35 month to month), for 28,000 credits a month. Refunds are typically granted within 14 days of payment if you've used no more than 1,200 credits.

Who it's for: a studio or a developer who wants to fine-tune or self-host a recent model, and a small team that wants a storyboard app on top. Skip it if your group makes more than $10M a year and you don't want a licensing negotiation.

The risk: the "Lite for business" trap. An agency that posts client ads made on Lite is outside the personal-use terms. And running LTX-2.5 yourself takes serious hardware: the checkpoint Lightricks publishes is a 22-billion-parameter model.

Wan is Alibaba's video model family. OpenArt now lists Wan 2.7 and Wan 3.0, and Higgsfield Wan 2.7, but the version to know if you want to own your pipeline is Wan2.2, because its license is plain: the models in its official repository are licensed under Apache 2.0, and Alibaba states it claims no rights over the content you generate.

Wan2.2 is the oldest model on this page (its weights came out in 2025), and it stays here for its license. It comes in several sizes. The 14-billion-parameter models handle text-to-video and image-to-video at 480p and 720p, and a speech-to-video model animates a character from an audio track. The 5-billion-parameter TI2V-5B model does text and image to video at 720p and 24 frames per second, and its single-GPU command runs on a card with at least 24 GB of video memory, such as an RTX 4090.

The price is your hardware and your time. The weights cost nothing to download; you pay for the GPU, owned or rented by the hour, and for the setup. For a product team that renders hundreds of short clips a month, a fixed GPU cost replaces a per-second bill. For one clip a week, a hosted model is simpler.

Who it's for: a developer, a studio with a GPU, or a product team that wants to build video generation into its own app. Skip it if you want a web interface and 1080p output: Wan2.2's models are listed at 480p and 720p.

The risk: version confusion. Before you build on a newer Wan version listed on a platform, find its weights and its license yourself: the Apache 2.0 license we read applies to the Wan2.2 repository. Apache 2.0 also leaves you accountable for what you generate, which matters for real faces and brands.