ChatGPT
Claude
The best AI video generator in November 2026 depends on the shot. For a dialogue close-up where the voice has to come out of the same generation, start with Seedance 2.5. For text on signs, hands and following a written camera move, start with Kling 4.0. For single-subject action and physics, Seedance 2.5 again. For drafts at the lowest list price, Gemini Omni 1.2 Flash at 360p, Wan 3.0 at 480p or the new Seedance 2.5 Flash. For a controlled 20-second move that drops into a color pipeline, Luma Ray3.2.
Those picks come from vendor specs, list prices read on November 13, 2026, two public leaderboards, and third-party tests we name below. We did not run our own test, and we do not invent scores. This is the November update to our 2026 AI video generator rankings and the April 2026 leaderboard.
Try the shot-type picks side by side
Seedance 2.5, Kling 4.0 and MiniMax H3 run in one generator, with the credit price shown before you generate.
The November 2026 shot-type leaderboard
| Shot type | First pick | Strong alternative | Why, in one line | On Oakgen? |
|---|---|---|---|---|
| Dialogue close-up | Seedance 2.5 | MiniMax H3 | Native audio and the better lip sync in CellCog's test; H3 tied best on a spoken line for VdoBloom | Both, yes |
| Single-subject action and physics | Seedance 2.5 | Kling 4.0 | Only model to finish VdoBloom's parkour jump and glass shatter | Both, yes |
| Two-person action and fights | Kling 4.0 | Seedance 2.5 | Fuser saw Kling keep two bodies apart where Seedance merged them | Both, yes |
| Product shots with text | Kling 4.0 | MiniMax H3 | Won sign text and hands for VdoBloom; all three tied on a plain product orbit | Both, yes (Kling silent) |
| Long takes (15 to 30 s) | Seedance 2.5 | Kling 4.0, Wan 3.0 | 30 s per pass; CellCog's 30 s take was a near tie | Seedance and Kling, yes |
| Keyframed, post-ready move | Luma Ray3.2 | Kling 4.0 | Up to 16 keyframes, HDR and 16-bit EXR (vendor claim) | No |
| Stylized and animated | Seedance 2.5 or Kling 4.0 | Wan 3.0 | Both held a painterly style in CellCog's test; no dedicated anime test found | Seedance and Kling, yes |
| Cheapest drafts | Omni 1.2 Flash (360p), Wan 3.0 (480p) | Seedance 2.5 Flash (480p) | Lowest list prices per second of the nine | No (see the Oakgen section) |
"First pick" is our reading of the evidence below, not a score. Where a pick rests on one third-party test with one take per model, we say so.
Specs and list prices, November 13, 2026
Prices are per second of output at each vendor's own list price, before plan discounts. Third-party resellers charge different amounts, which is why "Kling is three times cheaper than Seedance" is true on fal but closer to 1.5 times at BytePlus list prices.
| Model | Maker, launch | Max length | Max resolution | Native audio | List price per second |
|---|---|---|---|---|---|
| Seedance 2.5 | ByteDance, July 31, 2026 | 30 s | 1080p | Yes | $0.103 (480p), $0.231 (720p), $0.569 (1080p) on BytePlus |
| Seedance 2.5 Flash | ByteDance, announced Nov 12, 2026 | 30 s (reported) | 720p | Yes (reported) | $0.07 (480p), $0.15 (720p) on BytePlus |
| Kling 4.0 | Kuaishou, October 2026 | 30 s | 4K | Yes, stereo | 1.0 unit (720p), 1.5 units (1080p); about $0.157 and $0.235 at fal list |
| Gemini Omni 1.2 Flash | Google, GA Nov 5, 2026 | 15 s | 4K with upsampling | Yes | About $0.07 (720p), $0.023 (360p) on the Gemini API |
| Wan 3.0 | Alibaba, August 2026 | 30 s | 1080p | Yes | $0.05 (480p), $0.10 (720p), $0.20 (1080p) |
| MiniMax H3 | MiniMax, July 31, 2026 | 15 s | 2K | Yes, stereo | $0.08 (768p), $0.13 (2K) |
| Runway Gen-4.5 | Runway | 10 s | 720p | None listed in Runway's spec table | $0.12 |
| Luma Ray3.2 | Luma, June 9, 2026 | 20 s | 1080p, HDR, EXR | No audio option in Luma's price table | 100 credits per 5 s at 720p SDR |
| Veo 3.1 | 8 s in Flow's Quality tier | 4K | Yes | $0.40 with audio (720p, 1080p); Fast $0.10 at 720p |
Sources, all read November 13, 2026: ByteDance Seed, BytePlus pricing, Kling 4.0 guide and Kling API pricing, Google's Omni 1.2 post and Gemini API pricing, Wan 3.0 on Model Studio, MiniMax pricing, Runway Gen-4.5 help, Luma Ray3.2 and Luma pricing, Google Cloud Veo pricing.
Notes on the rows:
- Seedance 2.5 Flash. BytePlus announced it on November 12 with results "close to flagship Seedance 2.5" at "nearly twice the generation speed", and said the API opens on ModelArk on November 15 (BytePlus on X). Those are vendor claims. The 30-second length and the 10 image, 5 video and 5 audio reference limits come from Volcano Engine's announcement as reported by Pandaily. BytePlus's price page lists no 1080p. Our Seedance 2.5 Flash vs Seedance 2.5 post has the detail.
- Kling 4.0. Kling bills in units. Its API price table lists 1.0 unit per second at 720p and 1.5 at 1080p with no input video, and the package we read sold units at about $0.154 each after a small discount. CellCog read fal's list price as $0.157 at 720p and $0.235 at 1080p on October 29.
- Luma Ray3.2. Luma prices in credits. Its Plus plan is $30 a month for 10,000 credits, so 100 credits per 5 seconds works out to about $0.06 a second at 720p on that plan. That is our arithmetic, not a Luma price.
- Veo 3.1. Google shut down the Veo 3.1 preview in the Gemini API on October 22, 2026 (Google AI Studio), and Flow will retire Veo 3.1 Fast and Quality on December 5, 2026 (Google Flow Help). It stays on Google Cloud.
What the public leaderboards say
Two public boards rank these models from blind human votes. They disagree, and neither ranks by shot type.
| Model | Artificial Analysis text-to-video | Artificial Analysis image-to-video | Arena text-to-video |
|---|---|---|---|
| Wan 3.0 | 1175, #1 | 1108, #11 | 1476, #8 |
| MiniMax H3 (768p) | 1161, #3 | 1133, #4 | 1456, #11 |
| Seedance 2.5 | 1155, #4 | 1161, #1 | 1479, #7 (720p) |
| Gemini Omni 1.2 Flash | Not in the rows we read | 1086, #14 | 1540, #1 (preliminary) |
| Veo 3.1 | 965, #24 | 955, #27 | 1363, #14 (with audio) |
| Runway Gen-4.5 | Not in the rows we read | Not in the rows we read | 1223, #29 |
| Kling 4.0, Ray3.2, Seedance 2.5 Flash | Not listed | Not listed | Not listed |
Sources: Artificial Analysis text-to-video and image-to-video, read November 13, 2026; Arena text-to-video, last updated November 8, 2026. The two boards use different Elo scales, so compare ranks within a column, not numbers across columns. Artificial Analysis shows confidence ranges of about plus or minus 13 to 15 points, so Wan 3.0, MiniMax H3 and Seedance 2.5 sit close together on text-to-video.
What we take from it: Wan 3.0, MiniMax H3 and Seedance 2.5 are the general-purpose top tier by voter preference; Seedance 2.5 is the stronger pick when you start from an image; Omni 1.2 Flash's Arena lead is preliminary. Kling 4.0 is missing from both boards, so its case rests on third-party tests.
How we built this leaderboard
We read every source above on November 13, 2026: vendor launch posts, docs and price pages, the two leaderboards, three published same-prompt tests (CellCog, Fuser, VdoBloom) and one creator cost test on YouTube (Andy Lo). We pulled Oakgen credit quotes from Oakgen's generator the same day, which spends nothing.
We did not generate any clips for this article. Every quality finding belongs to the team that ran it, and each of those teams sells access to at least one of the models. Most tests were one take per model, which tells you about tendencies, not certainties.
Disclosure: Oakgen is our product. It runs three of the nine models here. We say below which ones it does not run.
Best AI video model for dialogue close-ups
The deciding question is whether the voice comes out of the same generation.
- CellCog, October 29. In its lip-sync test, CellCog supplied a voiceover to both models. Seedance 2.5 takes audio directly; Kling 4.0 got a cloned voice instead. Neither played the track verbatim, but CellCog says "Seedance sounds closer and syncs its lips better", scoring it 41 to 36. CellCog makes its own videos on Seedance 2.5 and says so (CellCog).
- VdoBloom, October 27. All three models it tested spoke a cafe line word for word. Kling 4.0 and MiniMax H3 tied at 4.25 out of 5, Seedance 2.5 scored 4.0 (VdoBloom). VdoBloom sells all three.
- Google says Omni 1.2 Flash gave "character speech an upgrade". That is a vendor claim with no third-party dialogue test we could find.
Our pick: Seedance 2.5 for talking close-ups, MiniMax H3 when budget matters. If the line must match a recorded voiceover exactly, lay the real track over the picture afterwards.
Best model for action shots
Split action into one subject and two.
One subject, fast motion or physics. VdoBloom's parkour test: Seedance 2.5 was "the only model to do the whole beat". Its wine glass test: Seedance broke the glass convincingly, MiniMax H3's glass "mostly stays whole", and Kling 4.0 refused the prompt as sensitive.
Two people moving fast. Fuser's capoeira test went the other way. Kling 4.0's "limbs stayed attached to the right person", while in Seedance 2.5's clip "the two bodies overlapped into one shape" for about half a second (Fuser). ByteDance itself lists "the stability of scenes involving interactions among multiple subjects" as room for improvement in its launch post.
CellCog's strict review also favoured Kling 4.0 on a pancake flip's arc, and says Kling stays on its watch list "for physics and action shots". So the honest answer is: Seedance 2.5 for solo stunts and breaking things, Kling 4.0 for fights and contact.
Best model for product shots
Product shots need the label to read and the hands to look right.
- VdoBloom's product orbit test was a three-way tie. Its sign-text test and card-shuffle test went to Kling 4.0, which spelled the sign right and flipped the correct card.
- CellCog's coffee bag test went to Seedance 2.5, which lifted a full bag "with both hands" the way a person would.
- Kling 4.0's guide lists text rendering in nine languages plus emojis and logos (vendor claim).
Our pick: Kling 4.0 when text or hands are on screen, MiniMax H3 for cheap product drafts, Seedance 2.5 when weight and handling matter. For compositing into a graded edit, Luma says Ray3.2 offers native HDR and 16-bit EXR export, which none of the others list.
Run product shots on Kling 4.0
3 to 30 seconds at 720p or 1080p, the same per-second credit price at both, up to 10 reference images.
Best model for long takes
Four of the nine make 30 seconds in one pass: Seedance 2.5, Seedance 2.5 Flash, Kling 4.0 and Wan 3.0. Ray3.2 goes to 20 seconds; Omni 1.2 Flash and MiniMax H3 stop at 15; Runway Gen-4.5 at 10.
Evidence on quality is thin. CellCog's 30-second single take was "the closest test of the ten", 33 to 32 out of 70 for Seedance. In Fuser's 8-second four-beat take, Kling 4.0 kept the ferryman and the order of events, while Seedance 2.5 "kept the mood, the light change and the payoff, and lost the man".
Two practical notes. Kling 4.0 turns multi-shot on by default, so write "one continuous take" if you want one. For a long move you want to control frame by frame, Ray3.2's 16 keyframes per clip are a vendor-stated feature no other model here matches.
Best model for stylized and anime shots
We found no dedicated anime test of these models, so treat this section as weaker evidence.
- CellCog's painterly mascot test: "Both hold the painterly style in every frame", 41 to 40 for Seedance 2.5.
- Kling's guide lists styles including "cinematic, retro DV, animation, felt, fantasy, and cyberpunk" (vendor claim), and its own example prompt is Aardman-style clay stop-motion.
- Alibaba says Wan 3.0 covers realistic, animation, software UI and text styles (vendor claim).
Stylized work has one practical advantage: stylized characters are far less likely to trip the real-face filters that Seedance applies to reference images. That matters for trends like the End of Watch locker scene.
Cheapest models for drafts
At vendor list prices, cheapest first:
| Draft option | List price per second | Limit to know |
|---|---|---|
| Gemini Omni 1.2 Flash, 360p | About $0.023 | Upscale later; 15 s max |
| MiniMax H3-Lite, 480p | $0.025 | Lighter variant of H3 |
| Wan 3.0, 480p | $0.05 | Alibaba ran a limited-time discount at launch |
| Seedance 2.5 Flash, 480p | $0.07 | API from November 15 |
| MiniMax H3, 768p | $0.08 | 15 s max |
| Seedance 2.5, 480p | $0.103 | Full model, draft resolution |
In Google Flow, any user can generate clips up to 10 seconds with Omni 1.2 Flash at 1 credit per second at 360p.
A creator test puts the per-second price in context. Andy Lo tested Seedance 2.5, Kling 4.0, FLUX 3 Max, MiniMax H3 Max and Wan 3.0 on the same two micro drama shots, with prices from November 4 to 6. His video description says mixing models, "Kling for faces, MiniMax for simple shots", got him to "about $12 a minute", and compares a 20-episode season at "$1,150 vs $240" (Andy Lo, YouTube, November 7, 2026). The lesson is his, and it holds: price times retries is the number that matters. Our guide to cost per usable clip shows how to track it.
What changed since April
Our April 2026 leaderboard was built on Artificial Analysis Elo scores from a different set of models. Since then: Seedance 2.5 (July 31), MiniMax H3 (July 31, open weights August 3), Wan 3.0 (August), Kling 4.0 (October) and Gemini Omni 1.2 Flash (November 5) all launched; Seedance 2.5 Flash was announced on November 12; and the Veo 3.1 preview left the Gemini API on October 22. The ranking also changed shape. In April one Elo number was enough. In November the top models are within error bars of each other, so the shot type decides.
For a head-to-head of the three newest models, see Gemini Omni 1.2 Flash vs Kling 4.0 vs Seedance 2.5.
Which of these models run on Oakgen
| Model | On Oakgen, November 13, 2026 | Oakgen price (10 s) |
|---|---|---|
| Seedance 2.5 | Yes: text, image and reference modes, 4 to 30 s, native audio | 746 credits (480p), 1,599 (720p), 3,935 (1080p) |
| Seedance 2.5 Flash | No | |
| Kling 4.0 | Yes: text, image, reference and video edit, 3 to 30 s, 720p or 1080p, silent output | 531 credits at 720p or 1080p |
| MiniMax H3 and H3 Max | Yes: H3 reference mode from 102 credits; H3 Max text mode | H3 Max: 169 (480p), 271 (768p), 541 (1080p) |
| Gemini Omni Flash | Listed, but we do not recommend it on Oakgen right now. Use Google Flow or the Gemini API | |
| Wan 3.0 | No | |
| Runway Gen-4.5 | No | |
| Luma Ray3.2 | No | |
| Veo 3.1 | No |
Seedance 2.5 quotes are for a 9:16 text-to-video clip with audio; reference video clips change how the provider bills the job. Kling 4.0's app does more than Oakgen's version: Kuaishou lists stereo audio, 4K and up to 10 keyframes, while Oakgen's Kling 4.0 is silent and tops out at 1080p.
So on Oakgen today, the practical split is Seedance 2.5 for talking and action, Kling 4.0 for silent product and reference-locked shots, and MiniMax H3 for cheap drafts. For the other six models, use the maker's own app or API.
Mistakes to avoid
- Picking a model by one leaderboard number. The top three are inside each other's error bars.
- Comparing reseller prices to list prices. Check the vendor's own page.
- Drafting at final resolution. On Oakgen's quotes, 10 seconds of Seedance 2.5 at 1080p costs about five times the 480p take.
- Trusting one take. Every test we cite used one or few takes per model. Run three of your own.
- Ignoring refusals. Kling refused a harmless wine glass prompt; Seedance refused invented portraits. A refused take still costs time.
- Asking a silent model for dialogue. On Oakgen, Kling 4.0 has no audio.
FAQ
What is the best AI video generator in November 2026?
No single model wins every shot type. On the boards we read on November 13, 2026, Wan 3.0 led Artificial Analysis text-to-video, Seedance 2.5 led its image-to-video board, and Gemini Omni 1.2 Flash led Arena's text-to-video board with a preliminary score. Third-party tests split by shot: Seedance 2.5 for physics and single-subject action, Kling 4.0 for text, hands, reference lock and camera direction.
Which AI video model is best for dialogue close-ups?
Seedance 2.5, if you need the voice made in the same pass. CellCog's October 29 test found it synced lips to a supplied voiceover better than Kling 4.0, though neither copied the voiceover verbatim. VdoBloom's test had Kling 4.0 and MiniMax H3 tied on a spoken line. On Oakgen, Kling 4.0 is silent, so use Seedance 2.5 there for talking shots.
Which AI video model is cheapest for drafts?
At vendor list prices on November 13, 2026, Gemini Omni 1.2 Flash at 360p (about $0.023 a second on the Gemini API), Wan 3.0 at 480p ($0.05) and Seedance 2.5 Flash at 480p ($0.07, API from November 15) are the cheapest of the nine models here. On Oakgen, MiniMax H3 starts from 102 credits and Seedance 2.5 at 480p is 746 credits for 10 seconds.
Which model is best for long takes?
Seedance 2.5, Seedance 2.5 Flash, Kling 4.0 and Wan 3.0 all make up to 30 seconds in one generation. Luma Ray3.2 goes to 20 seconds with up to 16 keyframes. In CellCog's 30-second single-take test, Seedance 2.5 and Kling 4.0 scored 33 and 32 out of 70, its closest test.
Is Veo 3.1 still worth using?
For most new work, no. Google shut down the Veo 3.1 preview in the Gemini API on October 22, 2026, and Google Flow will retire Veo 3.1 Fast and Quality on December 5, 2026. Veo 3.1 remains on Google Cloud at $0.40 a second with audio, and it ranked well below the newer models on both boards we read.
Did Oakgen test these models for this leaderboard?
No. We did not generate clips for this article. Specs and prices come from vendor pages, Oakgen prices come from Oakgen's generator quotes, and quality findings belong to the third parties named, with their methods and conflicts.
Which of these models can I use on Oakgen?
As of November 13, 2026, Oakgen runs Seedance 2.5, Kling 4.0 (silent, up to 1080p) and MiniMax H3 and H3 Max. It does not offer Seedance 2.5 Flash, Wan 3.0, Runway Gen-4.5, Luma Ray3.2 or Veo 3.1. Oakgen lists Gemini Omni Flash, but we do not recommend it on Oakgen right now; use Google Flow or the Gemini API for Omni.
Pick the model by the shot
Draft on MiniMax H3 or Seedance 2.5 at 480p, then finish on the model that fits the shot, all from one credit balance.
Sources
- ByteDance Seed, Introducing Seedance 2.5 (2026-07-31, read 2026-11-13)
- BytePlus ModelArk, model pricing for Seedance 2.5 and Seedance 2.5 Flash (read 2026-11-13)
- BytePlus on X, Introducing Dreamina Seedance 2.5 Flash (2026-11-12, read 2026-11-13)
- Pandaily, ByteDance Releases Seedance 2.5 Flash (2026-11-12, read 2026-11-13)
- Kling AI, Kling 4.0 User Guide (2026-11-04, read 2026-11-13)
- Kling AI, API pricing (read 2026-11-13)
- Google, Gemini Omni 1.2 Flash: fast and cost-effective with longer clips (2026-11-05, read 2026-11-13)
- Google AI Studio, Gemini Developer API pricing (read 2026-11-13)
- Google Flow Help, Manage your Google Flow credits (read 2026-11-13)
- Alibaba Cloud Model Studio, Wan3.0: 30-Second AI Video Generation from Any Input (2026-08-06, read 2026-11-13)
- Alibaba Cloud, Video generation model Wan3.0 officially launched (2026-08-24, read 2026-11-13)
- MiniMax, MiniMax H3 launch post (2026-07-31, read 2026-11-13)
- MiniMax, MiniMax H3 is now open source (2026-08-03, read 2026-11-13)
- MiniMax API docs, pay-as-you-go pricing (read 2026-11-13)
- Runway Help, Creating with Gen-4.5 (read 2026-11-13)
- Runway Dev, API pricing and costs (read 2026-11-13)
- Luma, Introducing Ray3.2 Model and API (2026-06-09, read 2026-11-13)
- Luma, plans and pricing (read 2026-11-13)
- Google Cloud, Agent Platform generative AI pricing for Veo 3.1 (read 2026-11-13)
- Google AI Studio, Generate videos with Veo 3.1, preview shutdown notice (read 2026-11-13)
- Artificial Analysis, AA-Video-T2V v2.1 leaderboard (read 2026-11-13)
- Artificial Analysis, AA-Video-I2V v2.1 leaderboard (read 2026-11-13)
- Arena, Text-to-Video leaderboard (updated 2026-11-08, read 2026-11-13)
- CellCog, Seedance 2.5 vs Kling 4.0: We Ran 10 Head-to-Head Tests (2026-10-29, read 2026-11-13)
- Fuser, Kling 4.0 vs Seedance 2.5: Same-Prompt Test (2026-10-23 test, read 2026-11-13)
- VdoBloom, Kling 4.0 vs Seedance 2.5 vs MiniMax H3: 8 Same-Prompt Tests (2026-10-27, read 2026-11-13)
- YouTube, Andy Lo, Make AI Micro Dramas 80% Cheaper (5 AI Video Models Tested) (2026-11-07, read 2026-11-13)
