Updated September 23, 2026. The short answer, by job:
- Best quality right now: Gemini Omni Flash (Google). It is first on both big blind-vote leaderboards and costs about $0.10 per second at 720p.
- Best value with sound: Wan 3.0 (Alibaba, about $0.08/s at 720P) and MiniMax H3 ($0.08/s at 768P). Both are in the top four on Artificial Analysis's audio leaderboard. Wan 3.0 clips run up to 30 s, MiniMax H3 up to 15 s.
- Animating a product photo: Seedance 2.0 and Kling 3.0. Both animate an approved still with start/end-frame control and cost $0.13–0.15/s at 720p with audio. Test them with your own photo.
- Longest single shot: Seedance 2.5 and Wan 3.0, both up to 30 seconds per generation.
- Avoid: Sora 2. OpenAI removes it from the API on September 24, 2026.
Want to see what these models do with your product rather than a demo reel? Upload one product photo, describe the shot, and get a clip back in Teamday's video studio.
Make your first video from a product photo →Real Work First: A 30-Second Film Made in Teamday
This is a finished, assembled video, not a single best-case clip. A second example, Dawn Delivery (15 s), was built from generated source frames, image-to-video, music and assembly. You can read the director's note and production inputs or watch the 15-second cut. The whole workflow is described on the short-form video job page.
September 2026 Comparison: Length, Audio, Price per Second
Prices are official API list prices checked on September 23, 2026, at 720p unless noted. "Audio" means the API can return sound generated with the video. Swipe the table sideways on mobile.
| Model | Maker | Max clip | Audio | Image-to-video | API $/sec (720p) | Best for | In Teamday |
|---|---|---|---|---|---|---|---|
Gemini Omni Flash (gemini-omni-1.1-flash) | 10 s (extend to 40 s) | Yes | Yes | ~$0.10 | Top quality per dollar, video editing | No | |
| Wan 3.0 | Alibaba | 30 s | Yes | Yes (first / last frame) | $0.083 (US/EU) · $0.10 (Singapore) | Long shots with sound on a budget | No (Wan 2.7 is) |
| MiniMax H3 | MiniMax | 15 s | Yes | Yes | $0.08 (768P) · $0.13 (2K) | Cheap high resolution, open weights | No |
| Seedance 2.0 | ByteDance | 15 s | Yes | Yes (first / last frame) | ~$0.15 | Product and people image-to-video | Yes |
| Seedance 2.5 | ByteDance | 30 s | Yes | Yes (up to 30 image references) | ~$0.23 | Long, reference-heavy scenes | Yes |
| Kling 3.0 | Kuaishou | 15 s | Optional | Yes (start / end frame) | $0.084 silent · $0.126 audio | Cinematic motion from a still | Yes (Kling 3 Pro) |
| Grok Imagine Video 1.5 | xAI | 15 s | Yes | Yes | $0.14 (+$0.01 per input image) | Fast short-form clips | Yes (720p) |
| HappyHorse 1.1 | Alibaba | 15 s | Yes | Yes | $0.124 (US) · $0.14 (Singapore) | Image-to-video in the Alibaba stack | 1.0 only |
| Veo 3.1 Standard / Fast / Lite | 8 s (extendable) | Yes, always on | Yes | $0.40 / $0.10 / $0.05 | 4K output (Standard, Fast) | No | |
| LTX-2.5 Fast / Pro | Lightricks | 20 s | Yes | Yes | $0.09 / $0.12 | Open weights, up to 4K | No |
| Runway Gen-4.5 | Runway | 10 s | No (API) | Yes | $0.12 | Hands-on creative editing | No |
| Luma Ray 3.2 | Luma | 10 s | No (API) | Yes (keyframes) | $0.06 · $0.24 at 1080p | HDR and EXR delivery for post-production | No |
| Sora 2 | OpenAI | — | — | — | Shuts down Sept 24, 2026 | Nothing new | No |
Not in the table: Midjourney Video V1 (no public API; image-to-video only, no audio), Adobe Firefly Video (no public per-second API price; the Firefly app also hosts partner models such as Veo, Kling, Runway and Luma), Pika 2.5 (API sold through a paid membership, no native audio) and Tencent HunyuanVideo 1.5 (open weights, 720p native, no official hosted API price). FLUX 3 Video (Black Forest Labs) and Meta Muse Video appear on the Arena leaderboard, but we could not verify public API pricing for them, so we do not rank them here.
What the Blind-Vote Leaderboards Say
Two public arenas rank video models by head-to-head human votes (checked September 23, 2026):
| Rank | Artificial Analysis, text-to-video with audio | Arena, text-to-video (updated Sept 21) |
|---|---|---|
| 1 | Gemini Omni Flash (1233) | Gemini Omni 1.1 Flash (1516) |
| 2 | Wan 3.0 (1229) | Gemini Omni Flash (1513) |
| 3 | MiniMax H3 Max (1227) | FLUX 3 Video (1493) |
| 4 | MiniMax H3 (1220) | Grok Imagine Video 1.5 agent (1492) |
| 5 | Seedance 2.0 720p (1210) | Seedance 2.0 720p (1479) |
Both boards put Kling 3.0 and Veo 3.1 outside the top ten, and Veo 3.1 costs four times as much as Omni Flash at 720p. That is the biggest change since July: Google's best video model is no longer Veo.
Leaderboards measure how good a single clip looks from a text prompt. They do not measure whether your product label survives, whether shot 3 matches shot 2, or what the finished ad costs. That is what the rest of this guide covers.
The Models, One by One
1. Gemini Omni Flash: the new quality leader
Google made Gemini Omni Flash generally available as gemini-omni-1.1-flash in August 2026. It generates 3–10 second clips at up to 4K and 24 fps, with a synchronized audio track by default. You can extend a clip 10 seconds at a time up to 40 seconds, and it also edits existing video. Pricing is per token: $17.50 per million video output tokens, and Google states 5,792 tokens per second at 720p, or about $0.10 per second. The preview alias gemini-omni-flash-preview is being retired, so build on the 1.1 ID.
Teamday: not integrated yet.
2. Wan 3.0: long clips with sound at the lowest price
Alibaba's Wan 3.0 (wan3.0-video) reached general availability in August 2026. It produces 2–30 second clips at up to 1080P with dialogue, music and sound effects, from text, a first frame, first and last frames, or references. Price depends on region: $0.083 per second at 720P in the US, Frankfurt, Tokyo and Hong Kong regions, and $0.10 in Singapore. A higher-tier Wan 3.0 Prime costs more. Unlike earlier Wan versions, Wan 3.0 has no open weights.
Teamday: runs the previous generation, Wan 2.7, for image-to-video at 720p. Wan 3.0 is not integrated yet.
3. MiniMax H3: 2K with audio for $0.13 a second
MiniMax replaced its Hailuo 2.3 line with MiniMax H3, released July 31, 2026. It generates up to 15 seconds with native stereo audio and supports image references. Official pricing is $0.08 per second at 768P and $0.13 at 2K, the cheapest 2K-with-audio rate we found. The weights are published on Hugging Face. Hailuo 2.3 is now listed as legacy, at $0.28 per 6-second 768P clip.
Teamday: not integrated.
4. Seedance 2.0 and 2.5: the image-to-video workhorse
ByteDance's Seedance 2.0 is still in the top five on both leaderboards. Seedance 2.5 (July 31, 2026) extends a single generation to 30 seconds and accepts up to 30 image references, so a product, a person and a location can stay consistent. Both generate audio with the video.
BytePlus ModelArk pricing is per token (width × height × fps × seconds / 1024), and audio does not change the price. For a 16:9 clip with no video input, that works out to:
- Seedance 2.0: about $0.07 per second at 480p, $0.15 at 720p and $0.37 at 1080p.
- Seedance 2.0 Fast: about $0.12 per second at 720p.
- Seedance 2.5: $0.23 per second at 720p and $0.57 at 1080p.
We covered the original launch in our Seedance 2 analysis.
Teamday: Seedance 2.0 and 2.0 Fast (4–15 s) and Seedance 2.5 (4–30 s, when your ModelArk account has access) run at 480p, 720p or 1080p. They accept a first frame and an optional last frame. Native audio is off by default so that voiceover and music can be mixed in assembly. Agents can turn it on for a clip.
5. Kling 3.0: cinematic motion from a still
Kuaishou launched Kling 3.0 in February 2026 and has since added Kling 3.0 Omni (references and editing) and Kling 3.0 Turbo. Clips run 3–15 seconds at up to 4K, with start and end frames. Kling's own price list is per second: $0.084 silent or $0.126 with audio at 720p, $0.112 or $0.168 at 1080p, and $0.42 at 4K. Test how well it keeps your product, logo area and first frame, not just how dramatic the motion looks.
Teamday: Kling 3 Pro text-to-video and image-to-video through FAL, 3–15 second clips, with audio optional. Kling 2.6 Pro remains available as a cheaper fallback.
6. Grok Imagine Video 1.5: fast short-form
xAI's Grok Imagine Video 1.5 became generally available in June 2026. At the end of July it gained text-to-video, reference images and native 1080p. It generates up to 15 seconds with audio in the same pass. The model page lists $0.08 per second at 480p, $0.14 at 720p and $0.25 at 1080p, plus $0.01 per input image. xAI's general pricing table shows only the $0.08 figure, so budget from the per-resolution rates.
Teamday: grok-imagine-video-1.5 for image-to-video and grok-imagine-video for text-to-video, 1–15 seconds at 720p, with the provider's audio.
7. HappyHorse 1.1: Alibaba's image-to-video tier
HappyHorse 1.1 (June 2026) generates 3–15 second clips at up to 1080P from text, an image or references. It costs $0.124 per second at 720P in the US region and $0.14 in Singapore. The 40% launch discount you may see quoted ended July 6, 2026.
Teamday: HappyHorse 1.0 image-to-video at 720p, 2–15 seconds. If you start from text only, Teamday generates the first frame for you, which adds an image cost.
8. Veo 3.1: still good, no longer the value pick
Veo 3.1 generates 4, 6 or 8 second clips (extendable) at up to 4K. Audio is always on, and it supports first and last frames plus up to three reference images. Pricing has three tiers:
- Standard: $0.40 per second at 720p or 1080p, and $0.60 at 4K.
- Fast: $0.10 at 720p.
- Lite: $0.05 at 720p, the cheapest audio-video rate in this guide.
Veo 3.1 Lite is a reasonable budget draft model. Veo 3.1 Standard is hard to justify now that Omni Flash ranks higher at a quarter of the price.
Teamday: not integrated.
9. LTX-2.5: open weights with audio
Lightricks' LTX-2.5 generates video and audio together at up to 4K. The API costs $0.09 per second (Fast, 720p) and $0.12 (Pro, 720p). The weights are open, which matters if you need to run video generation on your own infrastructure.
Teamday: not integrated.
10. Runway Gen-4.5 and Luma Ray 3.2: creative tools, silent APIs
Runway Gen-4.5 costs 12 credits per second at $0.01 per credit, so $0.12 per second. Its API returns 2–10 second clips at 720p with no audio option. Runway's value is its editing product. Its API also resells other models, including Veo 3.1, Seedance 2.5 and Gemini Omni Flash.
Luma Ray 3.2 (June 2026) is priced per 5-second clip: $0.30 at 720p and $1.20 at 1080p, or $0.06 and $0.24 per second. It adds HDR and 16-bit EXR export for post-production, and up to 64 keyframes. It also has no audio.
Teamday: neither is integrated.
What happened to OpenAI Sora?
OpenAI closed the Sora app on April 26, 2026 and removes sora-2, sora-2-pro and the Videos API on September 24, 2026. The deprecation page lists no replacement. Any "best video model" list that still recommends Sora is out of date.
Worked Example: What a 30-Second Ad Costs
A 30-second ad is typically four to six shots. The model cost is the price per second times the seconds you generate, not the seconds you keep. Below is list price for 30 seconds of 720p footage. The 3× column assumes you generate three takes for every second you use; that is an illustrative assumption, not a measured rate, so replace it with your own rejection rate.
| Model (720p) | $/sec | 30 s, one take each | 30 s, 3 takes each |
|---|---|---|---|
| Veo 3.1 Lite (audio) | $0.05 | $1.50 | $4.50 |
| Luma Ray 3.2 (silent) | $0.06 | $1.80 | $5.40 |
| MiniMax H3, 768P (audio) | $0.08 | $2.40 | $7.20 |
| Wan 3.0, US region (audio) | $0.083 | $2.48 | $7.43 |
| Gemini Omni Flash (audio) | ~$0.10 | $3.04 | $9.12 |
| Runway Gen-4.5 (silent) | $0.12 | $3.60 | $10.80 |
| Kling 3.0 (audio) | $0.126 | $3.78 | $11.34 |
| Grok Imagine Video 1.5 (audio) | $0.14 | $4.20 + $0.01 per input image | $12.60 + image fees |
| Seedance 2.0, 16:9 (audio) | ~$0.15 | $4.54 | $13.61 |
| Seedance 2.5, 16:9 (audio) | ~$0.23 | $6.93 | $20.80 |
| Veo 3.1 Standard (audio) | $0.40 | $12.00 | $36.00 |
Three conclusions:
- Generation is now the cheap part. Even with three takes per shot, most models cost less than $15 for a 30-second ad. Your time writing the brief, reviewing takes and fixing continuity costs more.
- Silent models are not cheaper if you then buy audio. Runway and Luma leave you to add sound separately. Omni Flash, Wan 3.0 and MiniMax H3 include it at a similar or lower price.
- Rejection rate matters more than list price. A $0.15/s model that keeps your product intact on the first take beats a $0.08/s model that needs five attempts.
How to Choose
Answer three questions in order:
- Do you start from a photo of a real product or person? If so, image-to-video fidelity comes first. Test Seedance 2.0, Kling 3.0 and Wan 3.0 with the same photo, and reject any take where the label, logo or face changes.
- Do you need sound in the clip? Omni Flash, Wan 3.0, MiniMax H3, Seedance, Kling (with audio), Grok and Veo generate it. If you will add voiceover and music in editing anyway, generate silent clips on Kling 3.0 at 720p for $0.084/s.
- How long is one shot? For shots over 15 seconds without cuts, only Seedance 2.5 and Wan 3.0 (30 s) and LTX-2.5 (20 s) do it in one generation. Otherwise, extend Omni Flash or Veo.
Then run a fair test. Split one 20–30 second concept into the same shot list, give every model identical prompts and source frames, and score each take:
| Criterion | What to measure |
|---|---|
| First-frame fidelity | Does the clip keep the approved composition, product and face? |
| Motion | Are bodies, objects, cloth, liquid and camera movement physically believable? |
| Continuity | Do identity, wardrobe, product, location and lighting survive across shots? |
| Prompt following | Did the requested action and camera move happen? |
| Audio | Are dialogue, ambience, effects and lip sync usable? |
| Rejection rate | Rejected takes divided by all takes |
| Cost per approved video | Model cost plus review and editing time, divided by approved deliverables |
Do not compare one model's best take with another's first take. Record every attempt.
Doing This in Teamday
Teamday's video studio and Reel, the AI video producer, handle the full job: brief, storyboard, source frames, clips, voice, music and final cut. You pick the model per shot from the routes Teamday supports today:
| Teamday route | Input | Clip length | Audio in Teamday |
|---|---|---|---|
| Seedance 2.0 / 2.0 Fast | Text, optional first and last frame | 4–15 s, 480p–1080p | Off by default; can be requested |
| Seedance 2.5 | Text, optional first and last frame | 4–30 s, 480p–1080p | Off by default; can be requested (needs ModelArk access) |
| Kling 3 Pro (FAL) | Text or image | 3–15 s | Optional |
| Kling 2.6 Pro, Wan 2.6 (FAL) | Text (Kling also image) | Short clips | Added in assembly |
| Grok Imagine Video / 1.5 (xAI) | Text / image | 1–15 s, 720p | Provider audio |
| HappyHorse 1.0, Wan 2.7 (Alibaba) | Image (auto-generated from text if needed) | 2–15 s, 720p | Added in assembly |
Every generated clip records the exact model ID it was made with, so when a better model ships you know which shots to regenerate.
Run the fair test on your own product instead of reading about it. Upload a product photo, and Reel storyboards the shots, generates takes on Seedance or Kling, and assembles the cut with voice and music for you to approve.
Animate your product photo in the video studio →Frequently Asked Questions
What is the best AI video generator in September 2026?
On blind-vote leaderboards, Google's Gemini Omni Flash is first on both Artificial Analysis (text-to-video with audio) and Arena, with Alibaba Wan 3.0, MiniMax H3 and ByteDance Seedance 2.0 close behind. For business video, Seedance 2.0 and Kling 3.0 remain strong image-to-video choices. The right pick depends on clip length, audio, resolution and price per second.
What is the cheapest good AI video model?
At official API list prices checked on September 23, 2026:
- Veo 3.1 Lite: $0.05 per second at 720p, with audio.
- MiniMax H3: $0.08 per second at 768P.
- Wan 3.0: about $0.08 per second at 720P (US region).
- Gemini Omni Flash: about $0.10 per second at 720p.
Thirty seconds of footage costs roughly $1.50 to $3 on these models, before retakes.
How much does a 30-second AI video ad cost to generate?
At 720p list prices, 30 seconds of generated footage costs about $1.50 (Veo 3.1 Lite) to $6.93 (Seedance 2.5), and $12 on Veo 3.1 Standard. If you generate three takes for every second you keep, multiply by three: roughly $4.50 to $21, or $36 on Veo 3.1 Standard. Voice, music, editing and review time come on top.
Which AI video models generate native audio?
Gemini Omni Flash, Veo 3.1, Wan 3.0, MiniMax H3, Seedance 2.0 and 2.5, Kling 3.0, Grok Imagine Video 1.5 and LTX-2.5 generate audio with the video. Runway Gen-4.5 and Luma Ray 3.2 expose no audio option in their APIs.
Is OpenAI Sora still available?
No. OpenAI closed the Sora app on April 26, 2026 and removes Sora 2 and the Videos API on September 24, 2026. Its deprecation page lists no replacement.
Which AI video models can I use in Teamday?
Teamday's video studio runs Seedance 2.0, Seedance 2.0 Fast and Seedance 2.5 (ByteDance ModelArk), Kling 3 Pro and Kling 2.6 Pro (via FAL), Grok Imagine Video and Grok Imagine Video 1.5 (xAI), HappyHorse 1.0 and Wan 2.7 image-to-video (Alibaba), and Wan 2.6 (via FAL). Gemini Omni Flash, Veo, Wan 3.0, MiniMax H3, Runway, Luma and LTX are not integrated yet.
Can I turn a product photo into a video?
Yes. All major models support image-to-video. In Teamday's video studio you upload a product photo, describe the motion, and it is used as the first frame of the clip.
Pick one product and one 30-second idea. Teamday's video studio takes you from photo to finished cut, and every clip records which model made it.
Start your first video →Related Guides
- Best AI image models 2026
- Best AI voice models 2026
- Best AI avatar models 2026
- AI video ads use case
- Teamday model catalog
Research date: September 23, 2026. Prices are official vendor API list prices from the pages linked above, in USD, without enterprise or promotional discounts. Artificial Analysis showed no update date; Arena was updated September 21. Model availability, regions and prices change often, so confirm the model ID, region, duration and audio setting before you commit a production budget.