Video · Comparison
Wan 2.6 vs Pika H3 Flex: Which 15-Second Multi-Shot Video Model Fits?
Pika H3 Flex suits a 15-second multi-shot clip built from many references, while Wan 2.6 suits a 15-second clip driven by text or one image at a flat per-second API price. Pika accepts up to 12 reference files and ships open weights. Wan 2.6 caps reference-to-video at 10 seconds, and Alibaba has not published its weights.
Both models were announced as 15-second, multi-shot, audio-capable video generators, and they are often searched together. This is a specification-and-price comparison from the makers' own documentation. We did not render side-by-side test clips, so it says nothing about which one looks better. It also covers where Ciyo, an AI design workspace with a conversational design agent and an infinite canvas for creating and editing images and videos, fits in.
How do Wan 2.6 and Pika H3 Flex compare on paper?
Both reach 15 seconds with sound and multiple shots, but they differ on references, resolution and licensing. Wan 2.6 is a hosted Alibaba model from December 2025. Pika H3 Flex is a post-trained version of MiniMax H3 released on October 29, 2026, with open adapters and a hosted API.
| Wan 2.6 | Pika H3 Flex | |
|---|---|---|
| Maker and release | Alibaba, announced December 16, 2025 | Pika, post-trained from MiniMax H3, released October 29, 2026 |
| Longest clip | 15 s for text-to-video and image-to-video; 10 s for reference-to-video | 5–15 s in every mode, including references |
| Multi-shot control | shot_type: multi parameter; for text-to-video it works only with prompt rewriting on | Shots and timings written into the prompt (Pika's examples use timecodes and [Shot n] labels) |
| References | Reference-to-video: up to 5 files (images 0–5, videos 0–3); text-to-video can take one audio file of 3–30 s | Up to 12 files: 9 images, 3 videos, 3 audio clips |
| Audio | Generated with the picture | Stereo, 32 kHz, generated with the picture; five separate stems (dialogue, score, SFX, Foley, ambience) via the API |
| Resolution | 720P and 1080P | Native 480p and 768p; 1080p, 2K and 4K come from a hosted upscaler |
| Aspect ratios | 16:9, 9:16, 1:1, 4:3, 3:4 | 11 ratios, from 21:9 to 9:21 |
| Weights | Not published (none in Wan-AI's Hugging Face organization on Nov 1, 2026) | Published as LoRA adapters plus MiniMax H3 base components |
| License | Hosted service terms | MiniMax H3 Community License; excludes the EU, UK, South Korea and the US |
Can Wan 2.6 make a 15-second multi-shot video with references?
Not in one pass. Alibaba's API reference lets wan2.6-t2v run 2 to 15 seconds, but wan2.6-r2v, the reference-to-video model, accepts only 2 to 10 seconds. So a 15-second Wan 2.6 clip is prompt-driven, with optional audio. A reference-driven Wan 2.6 clip stops at 10 seconds.
Wan 2.6's reference model takes at most five files, each holding one subject, and addresses them in the prompt as "character1", "character2" and so on. Reference videos are billed too: the usage record adds input-video seconds to output seconds. For one ten-second clip with a five-second reference video, you pay for fifteen seconds.
What does Pika H3 Flex do differently?
Pika H3 Flex takes up to 12 reference files in one request and still produces up to 15 seconds. The Ref2VA adapter accepts 9 images, 3 videos and 3 audio clips, and at least one image or video is required. Video and audio references must each be 2 to 15 seconds long, with 15 seconds combined per type.
Pika's showcase prompts, which are marketing examples and not benchmarks, direct a 15-second sequence shot by shot with timecodes such as 0–3 s and 3–5 s. The model card says Pika "post-trained MiniMax H3 to generate in six steps instead of 50". Pika claims a 15-second 768p clip in as little as seven seconds, but it defines that as transformer computation in a controlled benchmark, not end-to-end latency.
How much does a 15-second clip cost?
At list prices, a 15-second clip costs $1.50 at 720P or $2.25 at 1080P on Wan 2.6 text-to-video, and $0.60 at 768p or $1.05 at 1080p on Pika H3 Flex. Both bill per output second. Pika's launch promotion halves its prices for API Club members from October 29 to November 5, 2026.
Resolution is not like for like. Wan 2.6 renders at 720P or 1080P. Pika generates at 768p and upscales once in the same job, so its 1080p, 2K and 4K prices include that upsampling step.
| 15-second clip | Wan 2.6 | Pika H3 Flex |
|---|---|---|
| Text or image to video, 720P / 768p | $1.50 | $0.60 |
| Text or image to video, 1080P / 1080p | $2.25 | $1.05 |
| Text or image to video, 4K | Not offered | $2.40 |
| Reference to video, 720P / 768p | Not possible at 15 s; 10 s costs $1.00 plus input-video seconds | $1.80, plus input-video seconds |
| Ten clips at 720P / 768p | $15.00 | $6.00 |
Which one can you self-host or use commercially?
Only Pika H3 Flex gives you weights. Its adapters are Model Derivatives of MiniMax H3 under the MiniMax H3 Community License, and the repository bundles about 210 GB of base components. Pika says the package runs one request on one GPU and passed representative tests on an H100 80GB.
The license has limits that matter before you build on it. It excludes the European Union, the United Kingdom, the Republic of Korea and the United States. It also requires separate written authorization from MiniMax if your products generate more than $20 million in yearly revenue, and it requires the interface to display "MiniMax H3". The pages we read do not say whether Pika's hosted API carries the same territory limits, so check that with Pika and a lawyer if you operate in an excluded region. This is not legal advice.
"Open weights" here is narrower than it sounds. Wan 2.6 has no public weights at all: Wan-AI's Hugging Face organization lists Wan2.1, Wan2.2 and a few related models, but no Wan 2.5, 2.6 or 2.7 repository.
Is Wan 2.6 the newest Wan model?
No. Wan 2.6 launched on December 16, 2025, so pages that present it as new in late 2026 are repeating old news. Alibaba's Model Studio now also lists Wan 2.7 snapshots from April and June 2026 and Wan 3.0, which it describes as supporting four-modal reference input and up to 30 seconds of video.
Pika's own page shows a leaderboard snapshot with Wan 3.0 at 1324 and Pika H3 Flex at 1293, and Seedance 2.5 at 1256. Wan 2.6 does not appear in that snapshot. If you want the strongest Wan model, check Wan 3.0 instead of this one.
Which should you pick?
Pick Pika H3 Flex when you have more than five references, audio references, or need open weights. Pick Wan 2.6 when you work from text or one first frame, want 1080P rendered natively, or need a model with no territory exclusion in its license. Pick neither when you need 30 seconds. Use these rules as a starting filter, then test on your own shots.
Character-heavy ad or episode: Pika H3 Flex, because 9 image slots, 3 video slots and 3 audio slots cover several characters, a location and a voice.
Text-to-video story beat: Wan 2.6 at 15 seconds, with shot_type set to multi and prompt rewriting on.
Lowest list price per clip: Pika H3 Flex at 768p, from $0.60 per 15 seconds. Wan 2.6 is $1.50 at 720P.
Native 1080P: Wan 2.6. Pika's 1080p is an upscale of a 768p generation.
Self-hosting: Pika H3 Flex, subject to the license territory.
Where does Ciyo fit?
Ciyo's pricing page, read on November 1, 2026, lists neither Wan nor Pika. Its video models are Seedance 2.5, Kling 4.0, the Seedance 2.0 family and older Seedance versions. Ciyo is therefore a place to plan and compare shots, not a way to run these two models.
If your goal is a 15-second sequence, the closest listed option is Seedance 2.0: four to fifteen seconds per clip, 480p to 4K, a native-audio toggle and up to nine image, three video and three audio references according to its tool page. It costs 48 to 1,950 credits per clip depending on settings, and $10 buys 1,000 credits. That mirrors Pika's reference limits, but we have not compared the output quality.
For multi-shot work you can also build the sequence yourself on the infinite canvas. In our own storyboard test, two silent 5-second Seedance 2.5 clips at 720p, each from a start and end frame in Frames mode, cost 197 credits each (about $1.97). The walkthrough is in storyboard to video, and how much AI video costs covers the pricing logic. You can start from the AI Video Generator.
- tool page
- infinite canvas
- storyboard to video
- how much AI video costs
- AI Video Generator
- Alibaba Unveils Wan2.6 Series
- Wan text-to-video API reference (2.1–2.6)
- Wan reference-to-video API reference (2.6)
- Model Studio model pricing
- newly released models
- Wan-AI on Hugging Face
- Pika H3 Flex page
- research write-up
- Pika-Labs/Pika-H3-Flex on Hugging Face
- Ciyo pricing
- Seedance 2.0 on Ciyo
- Ciyo
Frequently asked questions
Is Wan 2.6 open source?
No. Alibaba offers it through Model Studio and its Wan site, and we found no Wan 2.6 repository in Wan-AI's Hugging Face organization on November 1, 2026. Wan 2.1 and 2.2 are published; 2.5, 2.6 and 2.7 are not.
Can Wan 2.6 do multi-shot video with reference images?
Yes, but only up to 10 seconds. The reference-to-video model accepts shot_type: multi, up to five reference files and 2 to 10 seconds. For 15 seconds you use text-to-video or image-to-video.
How many reference files does Pika H3 Flex accept?
Up to 12: 9 images, 3 videos and 3 audio clips. At least one image or video is required, and each video or audio clip must be 2 to 15 seconds long.
How much does a 15-second Pika H3 Flex clip cost?
$0.60 at 768p and $1.05 at 1080p for text or image to video at regular prices, and $1.80 at 768p for reference to video, plus input-video seconds. Launch prices are half through November 5, 2026 for API Club members.
Can I use Pika H3 Flex commercially?
Under the MiniMax H3 Community License, only outside the EU, UK, South Korea and US, with written authorization above $20 million in yearly revenue and a visible "MiniMax H3" notice. Confirm the terms for your situation before shipping.
Can I run Wan 2.6 or Pika H3 Flex in Ciyo?
Not as of November 1, 2026: Ciyo's pricing page does not list either. It lists Seedance 2.5, Kling 4.0 and Seedance 2.0 models for video, and the exact credit cost shows before each run.
Continue in Ciyo
Plan your shots on one board, then compare video models side by side in Ciyo. The credit cost appears before every run.