Wan 3.0 Video API Lands on APIXO With Reference-to-Video Control

What happened

Alibaba's Wan 3.0 video model is now available through APIXO, a third-party model platform. Developers can call it by API, and creators can use it in a browser Playground on apixo.ai. APIXO's pricing starts at $0.07 per generated second.1 APIXO offers three generation modes: text-to-video, image-to-video and reference-to-video. The reference mode matters most, because it lets users steer a clip with existing assets instead of a text prompt alone.1 APIXO lists Wan 3.0 Video among its August 2026 additions, next to MiniMax H3, Seedance 2.5 and FLUX 3. It also sells LoRA variants, including Wan 3.0 Video LoRA and Wan 3.0 Video Pro LoRA.3

Taken alone, this is a small item: one more reseller is hosting one more model. It does show where AI video is going in 2026. The selling point has moved from realistic output to control over the output, and Alibaba's choice to keep Wan 3.0 closed has made API middlemen the main way outside developers can reach it.

The model underneath

Alibaba opened a public beta for Wan 3.0 on August 6, 2026, through Alibaba Cloud Model Studio and Qwen Cloud.1315 The formal release came on August 24. That was one day after Alibaba priced a HK$80 billion (about $10.2 billion) share placement and said all net proceeds would go to AI.1214 Several outlets read the timing as Alibaba showing investors a product for its spending. That reading fits the facts: the company had just reported quarterly capital expenditure of about 67.7 billion yuan alongside a steep fall in net profit.1416

The main technical claim is a 30-second clip in a single pass. Wan 2.7 topped out at 15 seconds, and most mainstream generators produce somewhere between a few seconds and that same limit.1517 Output runs at 30 frames per second in 480p, 720p or 1080p.12 Coverage of the launch keeps returning to one input type in particular: documents. Wan 3.0 can take PDFs, slide decks, spreadsheets and webpages and build a video from what they contain.1416 Morphic lists the supported formats as doc, xls, ppt, pdf, txt, key, pages, numbers and md, plus any webpage URL.13 Alibaba says the model has already been used for short dramas, film, advertising, tourism promotion and music videos since the beta opened.14 TNGlobal adds self-driving and robotics simulation to the list of intended uses.15

How APIXO packages reference-driven generation

APIXO's reference-to-video mode accepts images, video clips and audio, plus either one public file URL or one web link. The file and the link cannot be sent in the same request.1 A single request can carry up to ten image URLs, up to five audio URLs, and up to five reference videos. Each video must run 1–15 seconds, and the videos together cannot exceed 15 seconds.1 These caps line up with what other hosts report. Winbuzzer describes the same 10/5/5 structure with 15-second totals.12 Layer's Wan 3.0 Reference listing has the same numbers.11 Morphic's figure of "up to 20 references" is simply those caps added together.13

The other controls are aimed at production work. Image-to-video takes a first frame and, optionally, a last frame, which lets a clip move between two fixed visual states.1 Duration can be set to any whole number from 2 to 30 seconds. Setting it to -1 lets the model choose the length.1 Reference video counts toward the 30-second ceiling, so five seconds of input leaves at most 25 seconds of new footage.1 Alibaba's own API reference shows the same features through a single media array. Each entry is tagged as a reference image, reference video, first frame, last frame or file. Its examples include turning a PowerPoint file into a product ad and restyling or extending an existing clip.10

In the coverage, the main difference between APIXO and Alibaba is packaging, not features. Alibaba presents Wan 3.0 as one model that combines generation, editing, reference use and motion driving. Wan 2.7 had split those jobs across separate models.19 APIXO, Pixazo and Venice split that single model back into separate mode endpoints.176 Each approach has a cost. Alibaba's version is simpler to describe. The split versions are easier to wire into existing tools, at the price of more endpoints to manage.

The pricing gap is the real story for developers

APIXO charges per generated second at three tiers: $0.07 for 480p, $0.12 for 720p and $0.225 for 1080p. Text, image and reference modes all use the same rates.12 Reference-video seconds are billed at the output rate. Image, audio, file and link inputs cost nothing extra. A 10-second 1080p clip costs $2.25, and the same clip with five seconds of reference video costs $3.375.1 Sound is on by default and watermarking is off.1

Alibaba's own list prices are lower: $0.05, $0.10 and $0.20 per second for the same three tiers.1217 That works out to a reseller premium of about 40% at 480p, 20% at 720p and 12.5% at 1080p, so the markup is heaviest on cheap draft renders. Venice also lists Wan 3.0 "from $0.07".6 An APIXO model page once showed $0.075 per second, which suggests prices were still settling.4

Against Western competitors, Wan 3.0 is cheap even after the markup. Dataconomy compared Alibaba's $0.20 per second at 1080p, or $12 a minute, with $0.40 per second for Google's Veo 3.1 standard tier.17 The Next Web summed it up as Wan 3.0 undercutting Veo.16 For teams producing many social-video variations, the gap between Alibaba's direct rates and a reseller's rates may matter less than how easy each is to access.

Why the middlemen exist

In practice, access is APIXO's main selling point. Coverage of Alibaba's own API is inconsistent. Winbuzzer reported that the Model Studio API was still in preview after the formal launch, and that developers needed approved access and credentials matching a supported region.12 Let's Data Science found Alibaba's launch page calling the API generally available without an application, while the English API reference, updated the same day, labelled it preview.18 Shortly after the beta opened, Siray reported that Alibaba had only promised full API access "soon".19 The Next Web described access as staged and said a members-only consumer site was planned.16 Alibaba also documents asynchronous jobs that usually take one to five minutes, plus regional endpoint matching.18 Both add work to any integration.

In that situation, a reseller offering a playground, a simple REST schema and a single bill is useful, and the per-second premium is the cost of that convenience. Nerdbot's coverage sees APIXO's main contribution as letting users go from browser testing to product integration in the same place.2 My reading is that this is more significant than the reference features. Those features belong to Alibaba, and many hosts now resell them.67911

The open-weights retreat

The reseller market also exists because Wan 3.0 cannot be self-hosted. Several outlets confirm it is closed and API-only, with no weights on Hugging Face, GitHub or ModelScope.81219 Wan 2.2 is the last open-weight flagship in the series. Versions 2.5 through 2.7 were already commercial, API-only models.13 Dataconomy calls this a break from the open-source approach that built the Wan name.17 That account slightly overstates how sudden the shift was, but it is right about the direction. Developers who chose Wan because they could run it themselves are now buying through cloud and reseller channels like everyone else.

That situation has also produced misinformation. Morphic warned that claims of 4K output, 60-billion-parameter spec sheets and Apache-licensed weights for Wan 3.0 are fabricated.13 Siray traced a claimed "1.3B and 14B" Apache release to the specs of Wan 2.1.19

How good is it, really?

Evidence on quality is thin. When the beta launched, Wan 3.0 had no technical report, no model card and no Artificial Analysis Arena entry.19 By late September, Layer listed Wan 3.0 Reference with an Arena Elo of 1362, ranking it fourth of 82 scored video models on its platform.11 Alibaba itself says audio texture and the accuracy of on-screen text still need work.12 That second weakness matters for a model sold on turning text-heavy documents into video. Winbuzzer also cited testing by Atlas Cloud that found repeated unwanted cuts in outputs meant to be one continuous shot.12 If that holds up, it undercuts the 30-second single-pass claim.

Competition is close, too. Seedance 2.5 launched within days of Wan 3.0 with similar 30-second clips and up to 50 reference inputs. Wan 3.0's most distinctive feature remains document and webpage ingestion.13

The takeaway

Wan 3.0 on APIXO is a fair snapshot of video generation in 2026. Developers can now steer a clip with images, motion clips, voices, documents and webpages, and this kind of reference-driven control is becoming standard.115 Clip length has doubled and pricing is below Western rivals.17 Alibaba, though, has kept the weights closed and is opening its own API gradually. That leaves room for resellers to charge a 12–40% premium for easier access.121 Teams should test against their own source material before committing, because the cut-and-text-accuracy issues are exactly what a document-to-video pipeline would hit first.