{ "schema": "pointcast.ai-video/v1", "url": "https://pointcast.xyz/reviews/ai-video", "title": "The new video machines", "dek": "Kling 4 just dropped, Sora is gone, and three Chinese labs are crowding Google at the top. Fifteen AI video services checked this week: how long, how sharp, whether they talk, and what they cost.", "byline": "cc", "asOf": "2026-09-29", "hero": "/images/ai-video/hero.jpg", "heroAlt": "A pixel-dithered specimen sheet in black, cobalt and yellow: a vintage movie camera on a tripod, a blank clapperboard, a film strip with a sun, a wave and a palm leaf, a play button, a timeline with diamond keyframes, a stereo waveform between two speakers, an hourglass, a server rack, a magic wand over a monitor, a price tag, a film reel and a red lightning bolt.", "heroCredit": "Illustration: Astra (Codex) for PointCast, after Mike Hoydich’s specimen sheets", "method": "Desk review from makers’ announcements, docs and pricing pages, checked September 29, 2026, plus news coverage where a maker publishes no page. Rankings are the Artificial Analysis Video Arena blind-vote leaderboards as seen September 29. PointCast has not run a side-by-side test of these models, so there is no star rating. Where a number comes only from resellers or aggregators, the card says so.", "disclosure": "Every link goes straight to the maker. PointCast earns no commission and no maker paid for or reviewed placement. PointCast’s own art desk uses some of these tools (Astra runs on OpenAI; the Higgsfield connector is installed here), which is worth knowing and changes nothing below. Prices are US list prices seen September 29, 2026; AI video prices move monthly.", "summary": "Kling announced 4.0 on September 28. So far only a lighter Flash version (720p, up to 20 seconds) is open, and only to Ultra yearly members; the full model (30 seconds, up to 4K) is promised for October. The field has settled on 15-to-30-second clips with sound made in the same pass. Alibaba’s Wan 3.0, MiniMax’s H3 and ByteDance’s Seedance now crowd Google’s Gemini Omni Flash at the top of the blind-vote charts. OpenAI is out: the Sora app closed April 26 and its API shut down September 24.", "endnoteTitle": "Omni Flash to start. Seedance for long scenes. Kling 4 when it opens.", "verdict": "If you want one tool today, start with Google’s Gemini Omni Flash: it tops the with-sound chart, it’s in the Gemini app you may already pay for, and the API is about ten cents a second. For long, consistent scenes, Seedance 2.5 does 30 seconds with fifty references now; Kling 4 promises the same in October, and it’s the one to watch. If you edit for a living, Runway and Luma are still the better workbenches, even though neither leads on raw quality. If you want to own the machine, run MiniMax H3 or LTX-2.5 on your own GPU. And don’t start anything new on Sora.", "leaderboard": { "source": "Artificial Analysis Text-to-Video Arena, with audio", "seen": "2026-09-29", "url": "https://artificialanalysis.ai/video/leaderboard/text-to-video", "ranks": [ { "rank": 1, "model": "Gemini Omni Flash", "maker": "Google", "elo": 1233 }, { "rank": 2, "model": "Wan 3.0", "maker": "Alibaba", "elo": 1229 }, { "rank": 3, "model": "MiniMax H3 Max (tuned by fal)", "maker": "fal / MiniMax", "elo": 1227 }, { "rank": 4, "model": "MiniMax H3", "maker": "MiniMax", "elo": 1220 }, { "rank": 5, "model": "Seedance 2.0 (720p)", "maker": "ByteDance", "elo": 1210 }, { "rank": 6, "model": "MAGI-2 Preview", "maker": "Sand.ai", "elo": 1156 } ], "note": "Kling 4.0 and Seedance 2.5 are too new to be ranked. Without audio, Wan 3.0 leads (1335), just ahead of Omni Flash (1332); Veo 3.1 is sixth and Kling 3.0 seventh." }, "groups": [ { "id": "new", "label": "Just dropped", "need": "The August–September launches, and the one everyone’s asking about." }, { "id": "studio", "label": "Studio workbenches", "need": "Editing, control and a timeline, not just a prompt box." }, { "id": "apps", "label": "Apps & model shops", "need": "One subscription, many models, quick social clips." }, { "id": "open", "label": "Run it yourself", "need": "Open weights on your own GPU. No meter running." } ], "picks": [ { "id": "kling-4", "name": "Kling 4.0", "brand": "Kuaishou · Kling AI", "group": "new", "award": "The one that just dropped", "released": "2026-09-28", "status": "Early access: 4.0 Flash for Ultra yearly members only. Full model promised for October.", "confidence": "confirmed", "price": 10, "priceText": "From ~$10/mo", "priceNote": "Kling’s Standard plan, as listed by resellers; Ultra runs about $180 a month. Kling hasn’t priced 4.0 yet. Check kling.ai.", "length": "30 s (Flash: 20 s)", "resolution": "Up to 4K promised (Flash: 720p)", "audio": "Yes, stereo", "how": "kling.ai web and app. Kling 3.x is also on the API and inside Adobe Firefly, Runway and Higgsfield.", "review": "Kling 4.0 is built for directing a scene, not rolling the dice on a prompt. The full model promises 30-second clips at up to 4K with stereo sound, up to 10 keyframes to pin the story beats, and up to 15 references (images, clips, characters, voices) to keep a face or a product the same from shot to shot. You can also edit an existing clip without regenerating it. What you can actually use today is smaller: the Flash version, 720p and up to 20 seconds, open since September 28 only to Ultra yearly subscribers.", "whatsNew": "Announced September 28, 2026: single clips go from 15 to 30 seconds, with keyframes, references and native stereo audio.", "bestFor": "Long, multi-shot scenes you direct beat by beat", "caveat": "Almost nobody can use it yet. Flash is 720p and has no keyframes, and it’s gated to the top yearly plan until October. One news report says the full model launched September 28. It didn’t.", "caveatSource": "https://www.atlascloud.ai/blog/tips/kling-4-flash-vs-full", "verdict": "The one to watch. Wait for October unless you already pay for Ultra.", "url": "https://kling.ai/", "cta": "Go to Kling →", "source": "https://x.com/Kling_ai/status/2104596718067257458", "sourceLabel": "Kling’s announcement" }, { "id": "gemini-omni-flash", "name": "Gemini Omni Flash", "brand": "Google DeepMind", "group": "new", "award": "Top of the charts", "released": "2026-08-27", "status": "Live in the Gemini app, Flow, YouTube Shorts and the Gemini API", "confidence": "confirmed", "price": null, "priceText": "~$0.10/s API", "priceNote": "Gemini API: $17.50 per million video output tokens at 5,792 tokens a second of 720p video, or about ten cents a second. Also included in Gemini app plans.", "length": "3–10 s per clip (extends to ~40 s)", "resolution": "Up to 4K", "audio": "Yes", "how": "Gemini app, Google Flow, YouTube Create, and the Gemini API, where Google now recommends it as the default video model.", "review": "Google’s new default isn’t Veo. Omni Flash is a Gemini model that makes video the way you’d have a conversation: generate a clip, then ask for changes in plain words. It sits first on the Artificial Analysis blind-vote chart for video with sound, and because it lives in the Gemini app and YouTube, it’s the one most people already have. Base clips are short, but you can extend them.", "whatsNew": "Announced at I/O in May, generally available as Omni 1.1 Flash in late August, and now Google’s recommended default over Veo.", "bestFor": "Generate, then edit by talking to it", "caveat": "Short base clips. Editing speech is restricted, and complex motion and on-screen text are still weak spots. The late-August release date and the 40-second extension come from secondary reports.", "caveatSource": "https://deepmind.google/models/model-cards/gemini-omni-flash/", "verdict": "The best place to start this week, and probably already in your phone.", "url": "https://gemini.google.com/", "cta": "Open Gemini →", "source": "https://ai.google.dev/gemini-api/docs/pricing", "sourceLabel": "Gemini API pricing" }, { "id": "wan-3", "name": "Wan 3.0", "brand": "Alibaba · Tongyi Lab", "group": "new", "award": "Best picture without sound", "released": "2026-08-24", "status": "Live on Alibaba Cloud and the Qwen Cloud API, by application", "confidence": "confirmed", "price": null, "priceText": "$0.05–0.20/s API", "priceNote": "API: $0.05 a second at 480p, $0.10 at 720p, $0.20 at 1080p.", "length": "30 s", "resolution": "1080p", "audio": "Yes", "how": "Alibaba Cloud Model Studio and the Qwen Cloud API, after an application. A members-only consumer site at wan.video is planned.", "review": "Wan 3.0 is first on the Artificial Analysis chart for video without sound and second with it. Its odd strength is input: hand it a PDF, a slide deck, a spreadsheet or a web address and it will make a video about it. Clips run 30 seconds in one pass.", "whatsNew": "Public beta August 6, launched August 24: single clips doubled to 30 seconds, and documents and URLs are accepted as prompts.", "bestFor": "Turning documents and web pages into video", "caveat": "The weights are closed now, which ends the open Wan line people ran at home. API access means applying first. We found no official Alibaba page, so these details come from several news reports.", "caveatSource": "https://dataconomy.com/2026/08/24/alibaba-launches-wan30-a-30-second-ai-video-generation-model/", "verdict": "Top-tier picture at a low per-second price, if Alibaba lets you in.", "url": "https://www.alibabacloud.com/product/modelstudio", "cta": "Go to Model Studio →", "source": "https://technode.global/2026/08/10/chinas-alibaba-releases-wan3-0-ai-video-model-in-public-beta-with-30s-clips-multimodal-inputs/", "sourceLabel": "TechNode on Wan 3.0" }, { "id": "seedance-2-5", "name": "Seedance 2.5", "brand": "ByteDance · Dreamina", "group": "new", "award": "Longest scenes you can make today", "released": "2026-07-31", "status": "Live in Dreamina (CapCut) worldwide; hosted on Pika, Higgsfield and others", "confidence": "confirmed", "price": 15, "priceText": "From $15/mo", "priceNote": "Dreamina Basic; the first month was on promotion for $1.50 when we checked. Pricing varies by region.", "length": "30 s (3-minute mode in beta)", "resolution": "Up to 4K", "audio": "Yes", "how": "Dreamina on the web and in the CapCut family of apps. In China it’s in Jimeng and Doubao.", "review": "Seedance 2.5 does now what Kling 4 promises for October: 30-second continuous clips, up to 50 reference images and clips to keep characters the same, and local edits that fix one detail without regenerating the whole thing. A beta long mode stretches to three minutes. Its older sibling, Seedance 2.0, is fifth on the with-sound chart; 2.5 is too new to be ranked.", "whatsNew": "Launched July 31: 30-second clips, a beta 3-minute mode, 50 references and local editing.", "bestFor": "Keeping a character consistent across a long scene", "caveat": "The three-minute mode is beta. Plans and prices differ by country, and it’s a ByteDance product, with the data questions that brings.", "caveatSource": "https://dreamina.capcut.com/seedance/seedance-2-5", "verdict": "The long-scene pick you can actually use this week.", "url": "https://dreamina.capcut.com/", "cta": "Open Dreamina →", "source": "https://www.prnewswire.com/news-releases/dreamina-launches-seedance-2-5--the-tool-of-ai-video-generation-that-reduces-clip-stitching-visual-drift-and-rework-302841204.html", "sourceLabel": "Dreamina’s launch release" }, { "id": "minimax-h3", "name": "MiniMax H3 (Hailuo 3)", "brand": "MiniMax", "group": "new", "award": "Cheap 2K with a voice", "released": "2026-07-31", "status": "Live in the Hailuo app, on the MiniMax API, and as open weights", "confidence": "confirmed", "price": null, "priceText": "Price not published", "priceNote": "MiniMax says 2K output costs under a third of mainstream models but lists no dollar price. The weights are free to download.", "length": "15 s", "resolution": "2K", "audio": "Yes, stereo with lip-sync", "how": "Hailuo web and app, the MiniMax API, open weights on Hugging Face (with ComfyUI support), and partners including Luma, Pika and fal.", "review": "H3 takes text, images, video and audio in one prompt and returns 2K video with stereo sound, including dialogue that matches the lips. It’s fourth on the with-sound chart, and a version tuned by fal sits third. Unusually for a model this good, you can download it.", "whatsNew": "Shown at WAIC on July 17 and released July 31 with open weights.", "bestFor": "Talking-head clips and dialogue with sound", "caveat": "Clips top out at 15 seconds. Secondary sources say the open license allows commercial use only for companies under $20 million in revenue.", "caveatSource": "https://www.marktechpost.com/2026/08/01/minimax-releases-minimax-h3-an-omni-modal-video-model-that-generates-15-second-2k-clips-with-native-stereo-audio/", "verdict": "Near the top of the chart, and you can download it.", "url": "https://hailuoai.video/", "cta": "Open Hailuo →", "source": "https://www.minimax.io/blog/minimax-h3", "sourceLabel": "MiniMax’s H3 post" }, { "id": "runway", "name": "Runway Gen-4.5", "brand": "Runway", "group": "studio", "award": "The editor’s workbench", "released": "2025-12-01", "status": "Live on the web, the app and the API", "confidence": "confirmed", "price": 12, "priceText": "From $12/mo", "priceNote": "Standard plan billed annually: 625 credits, or about 52 seconds of Gen-4.5. Resellers list the API at about $0.12 a second.", "length": "Up to a minute, multi-shot", "resolution": "1080p", "audio": "Yes", "how": "runway.com on the web and app, the Runway API, and inside Adobe Firefly.", "review": "Runway isn’t the top-ranked model anymore, and it hasn’t shipped a new frontier model since Gen-4.5 in December. What it has is the best editing room: Aleph 2.0 edits existing footage, and since July a Media Router picks between Runway’s models and other companies’ by quality, speed or cost. It’s where you finish a video, not just where you generate one.", "whatsNew": "Aleph 2.0 editing in May; a Media Router that sends each job to Runway’s or a third party’s model, July 23.", "bestFor": "Editing and finishing, with many models in one place", "caveat": "No new generation model in ten months. The API price comes from a reseller.", "caveatSource": "https://techcrunch.com/2026/07/23/runway-bets-on-ai-model-routing-as-generative-media-gets-crowded/", "verdict": "Still the best workbench, even though its own model trails.", "url": "https://runway.com/", "cta": "Go to Runway →", "source": "https://runway.com/pricing", "sourceLabel": "Runway pricing" }, { "id": "luma-ray3-2", "name": "Luma Ray3.2", "brand": "Luma AI", "group": "studio", "award": "For the color grader", "released": "2026-06", "status": "Live in Luma’s app and API", "confidence": "confirmed", "price": 25, "priceText": "From $25/mo", "priceNote": "Plus plan billed annually ($30 monthly): 10,000 credits, or about 100 seconds of 1080p Ray3.2.", "length": "20 s", "resolution": "1080p, 16-bit HDR / EXR", "audio": "No (sound comes from ElevenLabs tools inside Luma)", "how": "lumalabs.ai, the Luma API, and inside Adobe Firefly.", "review": "Luma is still the only one handing out real HDR: 16-bit, exported as EXR files that drop into a professional color pipeline. Ray3.2 adds frame-level direction with up to 16 keyframes per clip. If your video ends up in DaVinci Resolve, this is the one built for you.", "whatsNew": "Ray3.2 adds 16 keyframes per clip and keeps HDR/EXR output. The old Dream Machine and Ray2 are retired.", "bestFor": "HDR footage for a real color grade", "caveat": "The model doesn’t make sound; Luma adds it with built-in ElevenLabs tools. Not near the top of the blind-vote charts.", "caveatSource": "https://lumalabs.ai/news/introducing-ray-3-2", "verdict": "The pro’s pick when the file matters more than the prompt.", "url": "https://lumalabs.ai/", "cta": "Go to Luma →", "source": "https://lumalabs.ai/pricing", "sourceLabel": "Luma pricing" }, { "id": "veo-3-1", "name": "Veo 3.1", "brand": "Google DeepMind", "group": "studio", "award": "Extend and frame-match", "released": "2025-10-15", "status": "Live on the Gemini API, Vertex AI, Flow and the Gemini app", "confidence": "confirmed", "price": null, "priceText": "$0.05–0.40/s API", "priceNote": "API: Veo 3.1 Lite $0.05 a second at 720p, Fast $0.10, Standard $0.40.", "length": "8 s (extends to 148 s)", "resolution": "Up to 4K (Lite: 1080p)", "audio": "Yes", "how": "Gemini API, Vertex AI, Google Flow, the Gemini app, and inside Adobe Firefly.", "review": "There’s no Veo 4. Google’s docs now point new work at Omni Flash and keep Veo 3.1 for two jobs it does best: extending a clip scene by scene, and matching a first and last frame you supply. The cheap Lite tier, added this spring, is the lowest per-second price from a major lab.", "whatsNew": "No new model. The cheaper Veo 3.1 Lite arrived in spring 2026.", "bestFor": "Long extensions and exact first/last frames", "caveat": "Eight-second base clips, and Google now steers new users to Omni Flash.", "caveatSource": "https://ai.google.dev/gemini-api/docs/veo", "verdict": "The dependable tool for stitching a longer piece, at five cents a second.", "url": "https://deepmind.google/models/veo/", "cta": "See Veo →", "source": "https://ai.google.dev/gemini-api/docs/pricing", "sourceLabel": "Gemini API pricing" }, { "id": "adobe-firefly", "name": "Adobe Firefly", "brand": "Adobe", "group": "studio", "award": "Inside Premiere", "released": "2026-04-15", "status": "Live in Firefly and across Creative Cloud", "confidence": "confirmed", "price": null, "priceText": "Credits; price not checked", "priceNote": "Firefly and Creative Cloud plans include generation credits; partner models cost more credits. We couldn’t load Adobe’s current plan page.", "length": "Depends on the model", "resolution": "Depends on the model", "audio": "Depends on the model", "how": "Firefly on the web and app, Premiere, Photoshop, Express and Firefly Boards.", "review": "Adobe’s own video model isn’t the draw. The draw is that Firefly now runs more than 30 other companies’ models, including Kling 3.0, Veo 3.1, Runway Gen-4.5 and Luma Ray3.14, from inside the apps editors already use, plus Adobe’s commercially safe model for work where the rights have to be clean.", "whatsNew": "April 2026: Kling 3.0 and 3.0 Omni joined 30-plus partner models, with a new AI assistant and Quick Cut auto-editing.", "bestFor": "Commercially safe footage without leaving Creative Cloud", "caveat": "Adobe’s own model trails the partners, and the partner models eat credits faster. No specs published for Adobe’s own model.", "caveatSource": "https://news.adobe.com/news/2026/04/adobe-new-creative-agent", "verdict": "If you already live in Premiere, it’s the shortest path.", "url": "https://www.adobe.com/products/firefly.html", "cta": "Go to Firefly →", "source": "https://blog.adobe.com/en/publish/2026/03/19/adobe-firefly-expands-video-image-creation-with-new-ai-capabilities-custom-models", "sourceLabel": "Adobe’s Firefly update" }, { "id": "higgsfield", "name": "Higgsfield", "brand": "Higgsfield AI", "group": "apps", "award": "The ad kit", "released": "2026-09-21", "status": "Live on the web and app; hosts Veo 3.1, Kling 3.0, Seedance and more", "confidence": "unconfirmed", "price": 19, "priceText": "From ~$19/mo", "priceNote": "Starter billed annually, 270 credits, per a competitor’s published comparison. Higgsfield’s own pricing page wouldn’t load for us.", "length": "Depends on the model", "resolution": "Depends on the model", "audio": "Depends on the model", "how": "higgsfield.ai on the web and app, and an MCP connector for AI assistants.", "review": "Higgsfield doesn’t make its own video model. It sells the templates around other people’s: ad formats, UGC-style clips, camera moves, and Soul ID for keeping one character’s look across videos. On September 21 it announced new video-ad tools for small businesses, built with OpenAI’s GPT-6 Astra.", "whatsNew": "September 21: small-business video-ad tools, per an OpenAI customer story.", "bestFor": "Ads and UGC-style clips with a consistent character", "caveat": "No model of its own, and the credit math across models is hard to follow. Prices unconfirmed.", "caveatSource": "https://openai.com/index/higgsfield-from-prompt-to-production-with-astra/", "verdict": "Good for ads if you like templates. Check the credit math first.", "url": "https://higgsfield.ai/", "cta": "Go to Higgsfield →", "source": "https://openai.com/index/higgsfield-from-prompt-to-production-with-astra/", "sourceLabel": "OpenAI customer story" }, { "id": "pika", "name": "Pika", "brand": "Pika Labs", "group": "apps", "award": "Cheapest way to try the top models", "released": null, "status": "Live on the web and a social iOS app; hosts Seedance 2.5, MiniMax H3, Wan 3.0 and Omni Flash", "confidence": "unconfirmed", "price": 8, "priceText": "From $8/mo", "priceNote": "Starter billed annually ($10 monthly), 900 credits.", "length": "Depends on the model", "resolution": "Up to 1080p", "audio": "Yes (Pika Soundtrack)", "how": "pika.art on the web, the Pika iOS app, and an MCP server.", "review": "Pika has quietly become a store for other labs’ models: Seedance 2.5, MiniMax H3, Wan 3.0 and Omni Flash all run inside it, next to its own Pika 2.5. Its new Soundtrack tool adds synced sound effects, voice and music to any clip. At eight dollars a month, it’s the cheapest way we found to try several of the chart leaders side by side.", "whatsNew": "Relaunched as one studio for video, image and sound, plus Pika Soundtrack.", "bestFor": "Trying several top models under one cheap plan", "caveat": "Pika’s own model trails the ones it resells. The Pika 2.5 release date and specs come only from aggregators.", "caveatSource": "https://pika.art/", "verdict": "The tasting menu. Start here if you can’t decide.", "url": "https://pika.art/", "cta": "Go to Pika →", "source": "https://pika.art/pricing", "sourceLabel": "Pika pricing" }, { "id": "midjourney-video", "name": "Midjourney Video (V1)", "brand": "Midjourney", "group": "apps", "award": "Moving paintings", "released": "2025-06-18", "status": "Live on midjourney.com; image-to-video only", "confidence": "confirmed", "price": 10, "priceText": "From $10/mo", "priceNote": "Basic plan, as listed by aggregators; Midjourney’s docs wouldn’t load. HD video needs Standard or higher.", "length": "Up to 21 s", "resolution": "480p (720p HD on Standard and up)", "audio": "No", "how": "midjourney.com on the web. Animates your Midjourney images.", "review": "Midjourney’s video model is still V1, from June 2025: it takes one of your Midjourney images and sets it moving for a few seconds at a time, up to 21 seconds. No text-to-video and no sound. It stays on the list because nothing else animates the Midjourney look quite like Midjourney does.", "whatsNew": "Nothing new: Midjourney’s updates through September 24 were all image and editor changes. (Some blogs say video shipped in April 2026. It didn’t.)", "bestFor": "Bringing your Midjourney stills to life", "caveat": "No audio, no text-to-video, low resolution, and more than a year old.", "caveatSource": "https://updates.midjourney.com/", "verdict": "For your own Midjourney art only. Everything else has moved past it.", "url": "https://www.midjourney.com/", "cta": "Go to Midjourney →", "source": "https://updates.midjourney.com/introducing-our-v1-video-model/", "sourceLabel": "Midjourney V1 video post" }, { "id": "grok-imagine", "name": "Grok Imagine 1.5", "brand": "xAI", "group": "apps", "award": "Fast, inside X", "released": "2026-06-03", "status": "Live in the Grok apps and the xAI API", "confidence": "unconfirmed", "price": null, "priceText": "Price not checked", "priceNote": "Included in Grok plans; the API price wasn’t confirmed.", "length": "10 s", "resolution": "720p", "audio": "Unconfirmed", "how": "grok.com/imagine, the Grok iOS and Android apps, and the xAI API.", "review": "xAI’s video model is quick: resellers clock about 25 seconds to render a six-second 720p clip. It works best from a still image and lives where Grok users already are.", "whatsNew": "Video 1.5 reached the API in preview June 3, and reportedly became generally available June 16.", "bestFor": "Fast image-to-video in the Grok app", "caveat": "720p and 10 seconds. Only the preview date and specs come from xAI; the rest is from aggregators.", "caveatSource": "https://x.ai/news/grok-imagine-1-5", "verdict": "Fine if you’re already in Grok. Not a reason to start.", "url": "https://grok.com/imagine", "cta": "Go to Grok Imagine →", "source": "https://x.ai/news/grok-imagine-1-5", "sourceLabel": "xAI’s announcement" }, { "id": "ltx-2-5", "name": "LTX-2.5", "brand": "LTX (spun out of Lightricks)", "group": "open", "award": "Fastest on your own GPU", "released": "2026-08-11", "status": "Live as open weights and a managed API", "confidence": "unconfirmed", "price": 0, "priceText": "Free under $10M revenue", "priceNote": "Open weights are free for organizations under $10M in annual revenue; bigger companies need a paid license. Your GPU bill is yours.", "length": "Not confirmed", "resolution": "Up to 4K (LTX-2.3 family spec)", "audio": "Yes", "how": "Open weights on Hugging Face with ComfyUI support, or the managed LTX API.", "review": "LTX is the quick one: about seven seconds to generate a clip on good hardware, with sound, and weights you can download. Version 2.5 adds a new decoder that keeps faces, text and fast motion sharp, and holds a character’s identity across shots.", "whatsNew": "August 11: a sharper decoder and multi-shot identity hold.", "bestFor": "Fast local generation with no per-second meter", "caveat": "You need a strong GPU. Details come from one news report and aggregators; we found no official 2.5 page.", "caveatSource": "https://cryptobriefing.com/ltx-2-5-ai-video-model-release/", "verdict": "The home-lab pick for speed.", "url": "https://ltx.video/", "cta": "Go to LTX →", "source": "https://website.ltx.video/blog/introducing-ltx-2", "sourceLabel": "LTX blog" }, { "id": "magi-2", "name": "MAGI-2 Preview", "brand": "Sand.ai", "group": "open", "award": "Truly open, with sound", "released": "2026-08-05", "status": "Live as open weights (Apache 2.0)", "confidence": "confirmed", "price": 0, "priceText": "Free (Apache 2.0)", "priceNote": "Open source under Apache 2.0. You pay only for compute.", "length": "10 s", "resolution": "1080p (with the refiner)", "audio": "Yes", "how": "Weights on Hugging Face, code on GitHub.", "review": "MAGI-2 is the highest-ranked fully open model on the with-sound chart, sixth overall. It makes video and sound together, and its Apache 2.0 license has none of the revenue caps attached to H3 or LTX. It’s also a 114-billion-parameter model, so ‘run it yourself’ means real hardware.", "whatsNew": "Open-sourced August 5.", "bestFor": "No-strings open video with sound", "caveat": "Ten seconds a clip, and a very large model to host.", "caveatSource": "https://huggingface.co/sand-ai/MAGI-2-preview", "verdict": "The open-source benchmark. Bring a big GPU.", "url": "https://huggingface.co/sand-ai/MAGI-2-preview", "cta": "Get the weights →", "source": "https://github.com/SandAI-org/MAGI-2-preview", "sourceLabel": "MAGI-2 on GitHub" } ], "watchlist": [ { "name": "OpenAI Sora", "status": "Closed", "note": "The Sora app closed April 26, 2026, and the Videos API and every sora-2 model shut down September 24. OpenAI named no replacement, and Sora account data is deleted. Don’t start anything new on it.", "source": "https://help.openai.com/en/articles/20001152-what-to-know-about-the-sora-discontinuation", "sourceLabel": "OpenAI help center" }, { "name": "Kling 4.0 (full model)", "status": "Due October", "note": "The 30-second, 4K, 10-keyframe version is promised for October 2026. Until then, only the 720p Flash version is open, to Ultra yearly members.", "source": "https://www.atlascloud.ai/blog/tips/kling-4-flash-vs-full", "sourceLabel": "Flash vs. full, explained" }, { "name": "Seedance 2.5 long mode", "status": "Beta", "note": "Up to three minutes in one generation. Promising, but it’s labeled beta for a reason.", "source": "https://dreamina.capcut.com/seedance/seedance-2-5", "sourceLabel": "Dreamina" }, { "name": "HappyHorse 1.0", "status": "Ranked, little known", "note": "An Alibaba-ATH model sitting fourth on the no-audio chart. We couldn’t find a public product page, so it isn’t a pick yet.", "source": "https://artificialanalysis.ai/video/leaderboard/text-to-video", "sourceLabel": "Artificial Analysis" } ], "affiliateLinks": false, "handsOn": false }