Model selection
MiniMax for speed, Seedance for a low-cost trial, Veo for overall results
MiniMax H3 Max suits fast previews; Seedance's first-month offer lowers the cost of trying it. For a balance of visuals and motion, start by testing Veo 3.1 Fast.
TL;DR The main findings
- If you're in a hurry: choose MiniMax H3 Max first. In the café test using the same prompt, its 10-second video took 33 seconds to arrive, making it the fastest of the four models listed below in that round. Seedance 2.5 took 246 seconds.
- Low price trial: Seedance's Dreamina is priced at $1.50 for the first month, the lowest among the paid trial prices listed in this article. Only eligible users qualify; renewal is $15/month. For ongoing monthly fees, the Google AI Plus plan listed here is $4.99 and Kling Standard is $8.80. See the table below for allowances and available models.
- Overall effect: prioritize testing Veo 3.1 Fast. This recommendation considers both visual quality and how fully the action is performed, based on the coastal and rainy-night tests below. For atmosphere and individual frames, Seedance 2.5 is also worth trying; if camera movement matters most, compare Kling as well.
- Real-person references: check the restrictions of the Seedance service you plan to use. The official ModelArk does not accept direct uploads of images or videos containing real faces; there is a separate process for authorizing character materials. You cannot equate "generating characters" with "able to directly receive original actor footage."
The choices here are straightforward: start with MiniMax when you need to see a direction quickly; check Dreamina’s first-month offer when you want to try a round for less; include Veo when comparing results for a finished video. Below are the prices, timings and original clips behind these recommendations. The samples are limited, but they reveal specific differences: whether the latte art is completed, whether the camera moves as requested, and whether the bow touches the strings.
How much does an official website membership cost per month?
The table compares subscriptions to the official creation products, with common tiers priced in US dollars on a monthly billing basis. Kling uses its official website, MiniMax uses Hailuo, Veo uses Google AI/Flow, and Seedance 2.5 uses Dreamina. Prices were checked on October 18, 2026. First-month offers, renewals and the monthly equivalent of annual plans are treated separately.
| Model and official service | Monthly fee | Monthly quota | Before you subscribe |
|---|---|---|---|
| Kling / Kling | Standard: $6.99 for the first month, then $8.80 Pro: $25.99 for the first month, then $32.56 | Standard: 660 credits Pro: 3,000 credits | The page also lists regular prices of $10/$37. The figures above are the monthly subscription offers available on the date checked. The first-purchase offer applies only once. |
| MiniMax／ Hailuo | Standard: $10.50 on the date checked; regular price $14.99 Pro: $38 on the date checked; regular price $54.99 | Standard: 1,000 shells Pro: 4,500 shells | Monthly billing was selected. Use the regular price when budgeting beyond the discount period. Annual plans show a lower monthly equivalent but require a full year’s payment upfront. |
| Veo's platform/ Google AI／Flow | Plus:$4.99 Pro:$19.99 Ultra 5x:$99.99 Ultra 20x:$199.99 | Additional monthly Flow credits: 200 / 1,000 / 10,000 / 25,000, respectively | US official website prices. Plus has the lowest monthly subscription fee. The current plan page emphasizes Omni, so this does not establish that Plus includes the Veo 3.1 Fast tested here. Check your account’s model menu before subscribing. |
| Seedance 2.5／ Dreamina | Basic: $15 renewal Eligible first month offer: $1.50 | Example USD plan: 1,575 credits | The offer runs October 10–23 and is subject to region and payment-method restrictions. These prices are for the international Dreamina service, not the Chinese Jimeng service or Jianying/CapCut subscriptions. |
| Runway Gen-4.5／ Runway | Standard:$15 Pro:$35 Max:$95 | 625 / 2,250 / 9,500 credits, respectively | The website often defaults to the monthly equivalent of annual plans: $12/$28/$76. Those are not monthly billing prices. |
Prices come from the official links in each row. For Dreamina’s USD plans and offer conditions, see CapCut’s official pricing explanation (recorded October 16) . Currencies, credit bundles and taxes may vary by region. Your account’s checkout page determines the actual subscription charge.
Google's Flow credit guide also lists 50 daily credits, with subscription monthly allowances in addition to these. This comparison concerns the cost of subscribing; the lowest monthly fee does not necessarily mean the lowest cost per video.
Credits cannot be compared directly across services. Hailuo’s 1,000 shells, Kling’s 660 credits and Flow’s 1,000 credits each buy different amounts of generation. Choose the model and resolution first, then check the charge per generation to see whether the monthly allowance will cover your needs.
Speed: how long the community tests actually took
MiniMax's speed can be referenced in specific tests. On October 4, 2026, the Not an AI App tested multiple models using the same café prompt, each generating only once. Below are four models, timed from task submission to result receipt; Queries were made every 15 seconds, so the recorded value could be about 15 seconds longer than the actual completion time.
| Model | Actual output | Waiting time | Service and test conditions |
|---|---|---|---|
| MiniMax H3 Max Labeled Hailuo 3 Max in the test | 10 seconds, 768p | 33 seconds | OpenRouter; audio disabled in the request, but the returned file still contained an audio track |
| Veo 3.1 Fast | 8 seconds, 720p | 80 seconds | OpenRouter; audio enabled |
| Seedance 2.5 | 10 seconds, 720p | 246 seconds, approximately 4 minutes 6 seconds | OpenRouter; audio enabled |
| Kling 3.0 Pro | 10 seconds, 1080p | 293 seconds, approximately 4 minutes 53 seconds | OpenRouter; 720p requested, 1080p returned, audio enabled |
Source: Not an AI App’s same-prompt test, with original clips and timing methodology . These results apply to this test. Output specifications differ, so they are neither a waiting-time guarantee for official subscribers nor a fixed ranking of every model version.
Versions differ substantially. In another test by Krater, MiniMax H3 took 95.3 seconds to generate a 5-second clip at 2544 × 1456. The median for two 4-second, 720p Veo 3.1 Fast clips was 47.5 seconds; Kling 3.0’s was 52.3 seconds. H3 Max’s 33-second result cannot be applied directly to high-resolution H3 generation.
Source: Krater's October 2026 test , at 16:9 with audio disabled. Only one publicly timed H3 sample was available. Krater also provides generation services; this article cites its records without adopting its overall ranking.
The long wait in Seedance often includes queuing
Third-party service SeedVideo AI reported 724 completed tasks from September 15 to October 14. Seedance 2.5’s median wait was 4 minutes 42 seconds, and 90% completed within 11 minutes 23 seconds. The median was 4 minutes 38 seconds at 720p and 12 minutes 2 seconds at 1080p. These figures include provider queuing and file saving, but exclude failed tasks, rejections and unlimited-generation queues.
Source: SeedVideo AI task statistics and exclusion items . These are self-reported figures from a third-party platform, not an official ByteDance test, and this article has not independently audited them.
Community reports of waiting “several hours” more often concern unlimited-generation tiers on particular platforms. One user comparison, for example, reported waits exceeding 5 hours in TopView’s Seedance 2.5 unlimited queue. That suggests this plan can be slow; it does not establish that every Seedance generation requires 5 hours of computation. When facing a deadline, clarify the platform and the difference between its credit-based and unlimited modes, rather than simply asking whether the model is fast.
User Report: Reddit's Seedance 2.5 Unlimited Package Review . This is one person’s experience, without a standardized sample or a service-time guarantee.
A Runway timing is also available: on October 3, Quantslant waited 84 seconds for a 4-second, 720p Gen-4.5 rainy-night tram clip through OpenRouter. The subject and test date differ, so it is listed separately for reference.
Source: Runway test of Quantslant .
Why the same prompt can produce such different visuals
The model fills in details that the prompt leaves open. Whether a room has an extra column, how warm the lighting feels, whether the person remains in frame or the video cuts to a close-up—all these choices combine into the style you see. Style also affects suitability: product demonstrations need clearly visible objects, while atmospheric shorts may benefit from stronger light and shadow.
Start with two pairs of original clips, four in total. The first pair uses the opening 6 seconds of each clip; the second uses the opening 4 seconds. Original speed and aspect ratio are preserved, with no color grading. Play them side by side or open full-screen to inspect details. All samples depict realistic scenes; animation and anime need separate tests.
01 · Even when shooting latte art, the camera moves differently
Shared brief: a barista pours leaf-shaped latte art as the camera pulls back to reveal the café by the window. Watch the action inside the cup and check for sudden cuts.
Warm close-up, the latte art hasn't finished yet, and in about 4.6 seconds it cuts to a wider frame.
Natural window light is more apparent, and the milk pour and camera pullback remain in one continuous shot.
Original clips: Not an AI App, same test on October 4, 2026 . Both complete clips are about 10 seconds long. This comparison concerns shot execution, not sharpness across different source resolutions. The visuals are generated, not live-action footage.
02 · Violinist on a rainy night: assess atmosphere and action separately
Shared brief: a violinist on a rainy street at night, with neon in the background and a slow camera push-in. Watch the rain detail and the bow’s position.
Water droplets on the violin, streaks of rain and contact between the bow and strings are clearer; visuals and action feel more coherent.
The blue-green night and warm lights add a more dramatic feel, but the rain is not obvious, and the position of the bow is problematic.
Original clips: Krater’s rainy-night violinist test . Both clips are about 4 seconds long, with audio disabled for the test. The Seedance source is only 480p; it should not be used to judge the model’s maximum quality. This article provides visual commentary and excerpts only.
| Model | The visual features shown in the sample | Issues exposed in the review |
|---|---|---|
| MiniMax H3 | In Fuser's vase test, H3 added arched niches and columns, making the set richer than required. The potter's scene kept the face, hands, and vase in the same frame. | The camera pushes toward the vase only slightly; the richer set also introduces elements that were not requested. |
| Kling 3.0 Pro | With the same vase prompt, the camera pushes in more noticeably. The potter clip starts with the hands, then cuts to a closer view, unlike H3’s continuous shot. | No face appears in the first few seconds. If you need a continuous shot, check both the multi-shot setting and the prompt. |
| Veo 3.1 Fast | Krater's coastal sample is closer to the prompt word description; In the rainy night violin sample, the bow touches the strings and rainwater details are more credible. | In another set of café tests, latte art was completed from the very first frame, missing the gradual formation process. |
| Seedance 2.5 | In the same coastal and rainy-night review, the reviewer preferred its waves, night composition and dramatic feel. Individual frames stand out. | The rainy-night violinist sample did not generate rain, and the bow was incorrectly positioned. One beautiful frame does not necessarily sustain a complete action. |
| Runway Gen-4.5 | In Quantslant's rainy night tram samples, the wet stone pavement, streetlight reflections, and tracks look very realistic. | The tram moves faster than requested, and the sign contains garbled text. Attractive materials and accurate action need separate assessment. |
Corresponding samples: Fuser’s H3/Kling same-prompt test (September 28; two H3 generations and one Kling generation per prompt) ; Krater's coast/rainy night test ; Quantslant tram test . For the Veo latte-art issue, see Not an AI App café test . The table paraphrases the reviewers’ observations. No additional samples were generated for this article.
Kling’s official guide also explains the multi-shot setting: disabling it defaults to single-shot generation. Record this setting when comparing models. Another easily missed detail in the H3 vase example: Fuser found that automatic prompt expansion described the camera movement as a slight, slow push-in. Even when user inputs match, the descriptions reaching the generation stage may differ. Record whether automatic prompt expansion is enabled in your own comparisons.
If you're making a coffee commercial, you can start with the same product image and have both models each take a close-up of pouring coffee. First, check if the cup, hand, and liquid are correct, then see which one has better lighting and color for the brand. After picking one, try out another set of shots—it's easier to maintain the overall look than changing models for every shot.
Choose the duration around what needs to happen in the shot
Waiting 4 minutes and generating 10 seconds are two different things. Below is the duration of a single output clip; The final one-minute clip can be cut from multiple short shots.
| Comparison version | Duration per generation | How to choose |
|---|---|---|
| MiniMax H3／H3 Max | H3: 4 to 15 seconds H3 Max: 5 to 15 seconds | Short spoken lines and product actions can be divided into short shots. Check the specific mode for reference-input and resolution support. |
| Kling 3.0 | 3 to 15 seconds | The official website lists version 4.0; the sample test referenced here is 3.0, so the results cannot be directly applied to 4.0. |
| Veo 3.1 | 4/6/8 seconds | A particular creation interface may offer only some of these durations. Split long dialogue into shorter lines first, so a sentence is not squeezed into the ending. |
| Seedance 2.5 | Up to 30 seconds | Try it for a more complete action sequence. The longer the shot, the more closely you should check for changes in faces, disappearing objects and discontinuous movements. |
| Runway Gen-4.5 | 2 to 10 seconds | Suitable for generating by shot and then entering the editing process. |
Specifications: Hailuo’s official model and subscription comparison ; Kling VIDEO 3.0 Guide ; Google Veo 3.1 specifications ; Dreamina Seedance 2.5 ; Runway Gen-4.5 Guide .
Reference assets also affect the choice. If you only have product photos, start with image-to-video. To reuse an actor’s movements, choose a mode that accepts video or motion references. Writing “follow the reference movements” in the prompt does not mean the reference video has been supplied.
Real-person reference moderation: check Seedance’s restrictions first
For actor remakes, talking-head videos or ads featuring people, moderation may determine whether production can begin before image quality becomes relevant. This is particularly important with Seedance: BytePlus Official Portrait Guide explicitly states that the 2.5 and 2.0 series do not support directly uploading reference images or videos containing real human faces. Do not assume that original footage of an actor will be accepted as a motion reference.
| Model / service | Moderation of real-person assets | Impact on production |
|---|---|---|
| Seedance 2.5／ModelArk | Clear restrictions on direct uploading of real faces, covering images and videos. The official team also provides authorized character materials, preset digital roles, and trusted model output. | For a specific actor, first confirm that the service supports the authorized-asset workflow. A subscription or a written prompt cannot replace this check. |
| MiniMax H3 Max/Masonry hands-on test | H3 Max accepted the same AI-generated photorealistic portrait that Seedance 2.5 rejected. MiniMax was easier to get started with in this example. | Worth testing first for portrait references. This was a single-image test, not evidence that all real-person videos or subjects will pass moderation. |
| Veo／Google Flow | Google’s official guidance describes safeguards for minors and uploaded photos of people; some requests will not produce a result. | Test the actual assets when using real-person references. In particular, do not assume that acceptance of ordinary adult samples extends to content featuring children. |
| Kling official website | The community guidelines outline restrictions on portraits, privacy, and infringement, and use both technical and manual review. | A character-reference feature still requires compliance with asset and subject rules. This article does not provide rejection rates from a matched test across models. |
Moderation sources: BytePlus Official Portrait Guide (Updated October 8) ; Masonry’s test using the same portrait input ; Google Flow User Policy ; Kling Community Specification . The comparison covers specific restrictions and observations, without assigning an unmeasured “moderation leniency” score.
Seedance can generate people. ModelArk’s authorization workflow lets people verify their identity and grant authorization before adding images, videos and other assets to a character library. Trusted outputs are subject to conditions including the same platform, the same account and original files from within 30 days. Do not assume these capabilities are available in Dreamina or every third-party service. The ability to generate people also does not mean arbitrary original footage of real people can be uploaded.
Before paying, try out a few of your own materials
Try out three common shots: one for character movement, one for close-ups of the product, and one for the style scene you care about most. Keep the results you like, and note the reasons for failures. Just looking at the showcase videos selected from the official website, it's hard to know how many times your footage will need to be redone.
Include unusable generations in your budget. If a clip costs 60 credits and it takes four attempts to get one usable shot, the actual cost is 240 credits. This is only a calculation example. Even if the final edit uses just 4 seconds, the earlier generations have already consumed the allowance.
Also check if the free credit can test the target model and whether there are watermarks in the download. When involving people, first confirm that the actual material passes the review, then arrange subsequent storyboards and filming.
In rikoko, how can you choose?
rikoko uses multi-model access and selects generation modes such as text-to-video, image-to-video, reference-to-video, and first-and-last-frame according to the task. By default, it uses MiniMax, mainly considering its controllability in the rikoko workflow: storyboards, reference materials, and the duration and aspect ratio of each shot can be managed in the same process; if a shot is inappropriate, adjustments and retakes are made specifically for that shot.
When a quick direction check is needed, MiniMax mode that meets material requirements is prioritized. For special camera movements or visual styles, choices are made from the connected models. Which models are connected and what references can be uploaded depend on the services actually available in the current workspace.
Official website membership fees are used to compare various creative products; when using rikoko MCP to call generation services, actual connected services and the current quotation must be checked, and official membership points cannot be regarded as already available quota. After selecting shots, modifications to subtitles, music, and editing rhythm should reuse existing materials as much as possible, leaving the generation budget for parts that truly need to be reshot.
Keep reading
Three AI video workflows compared: MCP, Agents and Skills
Compare how Agents, Skills and MCP fit into video production, with a rikoko example that adapts a reference ad into a cold brew coffee campaign.
13 video analysis models: Gemini leads the cited video understanding benchmarks
Gemini 4 Argon scores highest in the benchmarks cited here. Compare frame-based analysis with GPT and Claude, specialist models for events and retrieval, and rikoko's Gemini analysis workflow.
AI video scripts and storyboards: narrative, shot design and prompts
A five-step guide from creative brief and audiovisual script to storyboard and generation prompts, using a 30-second product video with six complete shot examples and review criteria.
Translated from our October 2026 editorial drafts. Original authors, source dates and references are preserved; prices and rankings reflect the cited editions.