Step 13 of 22
Video models
Generate video, with sound, from text, images and keyframes. Video is the most expensive output per second. Clip length, control and audio vary more than headline quality.
How to choose
- Storyboarded scenes: models with keyframes and references.
- Volume and cost control: open-weight models on rented GPUs.
- Finished edits rather than clips: a creative suite on top of the models.
What to look for
The questions that separate the options, in order of how much they matter.
01
Arena Elo on the right board: with or without audio
Artificial Analysis ranks video models by Elo from blind side-by-side votes, with separate text-to-video and image-to-video boards and a With Audio / No Audio toggle. Each Elo carries a ± 95% interval and a Range column; in October 2026 the top six text-to-video models with audio sit within about 17 points, a statistical tie. If you will add your own sound, use the silent board; if you start from a product shot or storyboard frame, use image-to-video. Kling 3.0 Pro at 1080p is fixed at 1000 as the reference point, so scores only compare within that board.
02
Price per second at your resolution
The board's API Pricing column is the cost of one minute of 1080p video on the maker's API at default settings. Real costs swing by resolution and tier: Veo 3.1 is $0.40 a second at 1080p while Veo 3.1 Fast is $0.12, and Wan 3.0 is $0.20. Draft or fast tiers are worth using for iteration, with the expensive tier kept for final renders.
03
Clip length and extension
Single generations range from 8 seconds (Veo 3.1) to 30 (Wan 3.0, Seedance 2.5, Kling 4.0). Longer shots are usually built by extension, which can drift in identity and lighting with each step. If you need continuous takes, test the maximum single-pass length rather than relying on chained extensions.
04
Native audio
Several models now generate sound, dialogue and music in the same pass (Veo 3.1, Seedance 2.5, Wan 3.0, MiniMax H3, LTX-2.5, FLUX 3). Lip-sync and timing are much better when audio is generated with the picture rather than added afterwards. If the model is silent, budget for a separate speech or music model.
05
Control: references, keyframes and editing
For anything beyond a one-off clip you need control: first and last frames, multiple keyframes (Luma Ray3.2 takes up to 16), reference images for characters and products, and prompt-based editing of existing footage. Seedance 2.5 accepts up to 30 images, 10 videos and 10 audio clips as references. Check whether these controls are in the API or only in the maker's app.
06
Resolution and professional formats
Many models generate natively at 720p or 1080p and reach 4K by upscaling; check which. Some APIs cap resolution below the app (Seedance 2.5's ModelArk API stops at 720p). For post-production, HDR, ProRes and EXR export (Luma Ray3.2, Runway Gen-4.5) save conversion and grading work.
07
Licence, region and open weights
Hosted models license outputs under their terms, but availability varies by region: Seedance 2.5's ModelArk API is not offered in the US, and MiniMax H3's open weights need an application for use in the US, EU, UK and South Korea. LTX-2.5 weights are free for companies under $10 million in revenue. Open video weights also need serious GPUs; MiniMax's reference deployment uses four.
How the options differ
Most video models are closed and hosted, sold per second through an API, through the maker's own app on a subscription, or both. Google, ByteDance, Alibaba, Kuaishou and Black Forest Labs lead the blind-vote boards, and several are also resold inside apps such as Runway, which bundles third-party models alongside its own Gen-4.5.
Open-weight video models (MiniMax H3 base, LTX-2.5) let you self-host, fine-tune with LoRAs and keep footage private, but their licences carry revenue or region conditions, parts of the pipeline may stay API-only, and they need multi-GPU or high-end hardware for good speed.
10 tools, compared
Ranked by our rating. Open any review for sources, prices and the full verdict.
| Tool | Rating | Our verdict | Best for | Price |
|---|---|---|---|---|
| Wan 3.0model | 4.5/5 | Wan 3.0 wins the blind votes by a hair and runs to 30 seconds a clip, but Alibaba has kept its weights and its technical details to itself. | Creators and developers who want leading video quality and long takes without premium pricing. | Qwen Cloud: $0.05/second at 480p, $0.10 at 720p, $0.20 at 1080p list ($12 per minute at 1080p), currently shown 30% off. Runway: same rates, or $0.068 to $0.28 for the wan3_prime variant. |
| Seedance 2.5model | 4.4/5 | Seedance 2.5 sells control over continuity, and charges nearly three times the board leader's per-minute price for it. | Ad and content teams who need the same characters, products and voices held steady across many shots. | ModelArk: $10.70 per 1M tokens without video input, $6.40 with video input (about $0.51 for 5 seconds at 480p, $1.16 at 720p). Runway: $0.20 to $0.68 per second by resolution, 80-credit minimum. |
| MiniMax H3 (Hailuo 3.0)model | 4.3/5 | Near-frontier video at a seventh of Seedance's price, though the open weights are off limits in the four markets most buyers sit in. | Teams that want leading video quality cheaply through an API, and teams outside the US, EU, UK and Korea that want to self-host or fine-tune a top video model. | MiniMax API: $0.08 per second at 768p, $0.13 at 2K ($4.80 per minute at 768p); 2K upscale of an existing clip $0.05 per second. Runway: $0.10 per second at 768p. Weights free under the MiniMax H3 Community License. |
| Gemini Omni Flash 1.1model | 4.3/5 | Gemini Omni Flash trades a few Elo points for the cheapest route to editing a clip by conversation, and with the Veo previews retired this week it is the only video model on the Gemini API price list. | Developers on the Gemini API who want to generate and revise short clips through conversation. | $1.50 per 1M input tokens and $17.50 per 1M video output tokens, about $0.10 per second of 720p video; no free tier. Runway: $0.034 to $0.30 per second by resolution. |
| FLUX 3model | 4.2/5 | A capable preview with real production controls, priced and rate-limited like a model still finding its feet. | Studios and agencies that iterate on shots in drafts and need 4K delivery from a vendor they already use for images. | Per second of video: Draft HD $0.06; HD $0.17, FHD $0.29, QHD $0.40, 4K $0.80 (continuation $0.12 to $0.95); audio included. FLUX 3 Image about $0.048 per image. |
| Kling 4.0model | 4.1/5 | Kling 4.0 has the longest spec sheet in video generation and, so far, nothing independent to back it up. | Creative teams already in the Kling app who want long, keyframe-directed shots and can wait for an API. | Kling app subscriptions; Kling 4.0 API pricing not published. Predecessor Kling O3 on fal costs $0.14 a second (Pro, with audio) and $0.42 a second at 4K. |
| Luma Ray3.2model | 3.9/5 | A grading-suite tool rather than a leaderboard model: worth its price for directable HDR shots, not for anything that needs sound. | VFX, post-production and agency teams who need directable shots that drop into an HDR grading pipeline. | API per clip: 5-second text/image-to-video $0.15 (540p), $0.30 (720p), $1.20 (1080p) in SDR; 10-second $0.45/$0.90/$3.60; HDR 2x, HDR+EXR 3x. App: Plus $30, Pro $90, Ultra $300 a month. |
| Runway Gen-4.5model | 3.8/5 | Gen-4.5's strongest remaining argument is HDR and ProRes export, and Runway's standard terms let it train on whatever you feed it. | Studios and post houses already on Runway that need generated shots delivered in ProRes or HDR. | Free tier with 125 one-time credits; Standard $15/month ($12 yearly), Pro $35 ($28), Max $95 ($76). API: 12 credits ($0.12) per second, plus $0.05 for ProRes and $0.20 to $0.40 for HDR. |
| LTX-2.5model | 3.8/5 | Easy to self-host, fine-tune and wire into ComfyUI, LTX-2.5 trades 200 Elo of picture quality for control. | Developers and smaller studios who want an open video model they can fine-tune and run on their own GPUs. | Weights free under $10M annual revenue (LTX-2.x Community Licence). API per second: Fast $0.09 (720p), $0.13 (1080p), $0.19 (1440p), $0.30 (4K); Pro $0.12, $0.17, $0.25, $0.39. |
| Veo 3.1model | 3.3/5 | Veo 3.1 is now an enterprise-cloud legacy model: still GA on Google Cloud, gone from the Gemini API, and 170 Elo behind its own replacement. | Teams with existing Veo pipelines that depend on long extensions or end-frame control and can run on Google Cloud. | Google Cloud, per second: Veo 3.1 $0.40 with audio or $0.20 without (720p/1080p), $0.60/$0.40 at 4K; Fast $0.10 to $0.12 with audio; Lite $0.05 to $0.08 with audio, from $0.03 without. |
Side by side
The specifications that matter for this step. Checked 2026-10-19; confirm on the vendor’s site before you commit.
| Tool | Provider | Weights / licence | Max clip | Max resolution | Native audio | API price |
|---|---|---|---|---|---|---|
| Wan 3.0 | Alibaba | Closed (Wan 2.2 was open) | 30 s | 1080p | Yes | $0.20/s (1080p) |
| Seedance 2.5 | ByteDance | Closed; API not in US | 30 s | 720p on ModelArk API | Yes | About $1.16 per 5 s (720p) |
| MiniMax H3 (Hailuo 3.0) | MiniMax | Open base; MiniMax licence | 15 s | 768p open; 2K via API | Yes, stereo | About $0.08/s (768p) |
| Gemini Omni Flash 1.1 | Closed | 10 s | 4K (upscaled) | Not published | About $0.10/s (720p) | |
| FLUX 3 | Black Forest Labs | Closed (open Dev announced) | 20 s | 4K | Yes | $0.29/s (FHD) |
| Kling 4.0 | Kuaishou | Closed | 30 s | 4K | Yes, stereo | API price not published |
| Luma Ray3.2 | Luma AI | Closed | 20 s | 1080p, HDR, EXR | Not published | $1.20 per 5 s (1080p) |
| Runway Gen-4.5 | Runway | Closed | Not published | Not published | Not published | $0.12/s |
| LTX-2.5 | Lightricks | Open; free under $10M revenue | 20 s (Fast API) | 4K (up to 10 s) | Yes | Fast $0.13/s (1080p) |
| Veo 3.1 | Google DeepMind | Closed | 8 s (extend to 148 s) | 4K | Yes | $0.40/s (1080p); Fast $0.12/s |
Head to head and ranked guides
Where this stage fits
- Video production: we would start with Veo 3.1, Kling 4.0, Wan 3.0.