Wan 3.0: Alibaba's Cheap, Fast Video Model Hits GA

Short version: Alibaba's Wan 3.0 went generally available on 24 August 2026. It generates 2 to 30 second clips at up to 1080p with audio on by default, and at $0.50 for a five-second 480p clip it is the cheapest credible way to find out whether an idea works before spending real money on it.

Disclosure: Leaxor is mine and now runs Wan 3.0.

What Wan 3.0 is

Wan is Alibaba's video model line. Version 3.0 arrived on fal on 24 August with text-to-video, image-to-video and reference-to-video endpoints, all marked generally available rather than preview.

fal publishes the endpoint and its rates in full. The specification is unremarkable in a good way. Durations from 2 to 30 seconds. Resolutions at 480p, 720p and 1080p. Aspect ratios covering 16:9, 9:16, 1:1, 4:3 and 3:4, plus an adaptive setting. Audio on by default. Nothing here is a headline, and everything here works.

Why GA matters more than the version number

This is the part worth understanding if you build on these models rather than just using them.

Wan 2.5 shipped as a preview endpoint. Preview endpoints are the ones providers retire, usually with little notice, and anything wired to one is a future outage waiting for a date. Wan 3.0 arriving as GA is the migration path, not just a newer number.

If you have a pipeline pointing at a Wan 2.5 preview URL today, moving it is maintenance you have already been assigned and have not been told about.

What actually changed

• Duration to 30 seconds, up from a much shorter ceiling. Also a smart-duration mode: leave the length unset and the model picks one from your prompt.
• Reference-to-video. A third input mode beyond text and a single start image, taking reference media to steer the result.
• A thinking toggle, off by default, which spends more compute reasoning about the prompt before rendering.

Alibaba also ships a Prime variant at roughly a third more per second, for when the base model's output is not holding together.

What Wan 3.0 costs

Priced per render on Leaxor:

That 480p row is the useful one. At $0.50 for five seconds you can try ten framings of an idea for $5.00, then render the winner once at 1080p.

The settings that actually matter

Four of Wan 3.0's parameters change the output enough to be worth understanding before you spend anything. The rest you can leave alone.

Duration, and the case for leaving it blank

Any whole number from 2 to 30 seconds. The interesting option is not setting one at all: leave duration unspecified and the model picks a length from your prompt and any reference media.

That is genuinely useful when the prompt describes an action with a natural end, like a door closing or a wave breaking. It is a bad idea when you are cutting to music and need exactly six seconds, because you will get seven and pay for it.

Resolution is the price dial, not the quality dial

480p, 720p and 1080p, and the price roughly doubles at each step. What does not change is composition, motion or how well the model understood you. A shot that falls apart at 480p falls apart at 1080p, more expensively.

So the resolution setting is really a question about which stage you are at, not how good you want the result to be.

Audio is on by default

Worth knowing, because it is unusual. Most models return silence. Wan 3.0 generates sound with the picture unless you turn it off, which means a clip you assumed was silent may arrive with ambience you did not plan for and have to mute in the edit.

Prompt expansion and thinking

Prompt expansion is on by default and rewrites your prompt before rendering. It generally helps short, vague prompts and gets in the way of long, precise ones. If you have written a careful three-clause description of a shot and the result ignores half of it, this is the first thing to switch off.

Thinking is off by default and spends more compute reasoning about the prompt first. It is worth enabling for prompts with several things happening at once, and pointless for a single subject on a plain background.

What to actually use it for

Three jobs where Wan 3.0 is the right call rather than merely an available one.

Finding the shot. This is the main one. At $0.50 you can generate ten framings of the same idea for $5.00, pick the one that works, and only then decide whether it deserves a more expensive model. Most people skip this stage and it is why their bills look the way they do.

Long single takes. Thirty seconds in one pass, with no seam to hide, is unusual. Most models cap between 8 and 15, which forces you to stitch and then deal with the join.

Anything with reference material. The reference-to-video mode takes media to steer the result rather than starting from a sentence, which is far more predictable when you already know what the thing should look like.

Where it is the wrong call: hero shots where quality carries the piece, and anything needing genuinely good synchronised audio, which is Flux 3 territory.

Where it sits against the rest

Wan 3.0 is not the best-looking model available and it does not need to be. It is the one whose price lets you be wrong repeatedly, which for most real work is more valuable than a marginal quality edge.

Against Flux 3, Wan is well under half the price at 720p and behind on audio. Against MiniMax H3 Max they match at 480p, and H3 Max caps at 768p. Against Gemini Omni Flash, Google goes cheaper at 360p and reaches 4K, which Wan does not.

Should you use it?

Yes as your default for drafting. Cheap, quick, coherent enough to judge a shot.

Yes if you need 30 seconds in one pass, which few models offer.

No if the clip is the hero shot and quality is the whole point. Spend more there.

No if you wanted a finished video rather than footage. That is a pipeline, not a model.

Wan 3.0, the Prime variant, and its image-to-video and reference-to-video modes are all live in Leaxor's clip generator, priced per render with credits that do not expire and no subscription required. The rest of this month's launches are in the August 2026 roundup, and every model's per-second price is in the cost breakdown.

Reading this with an AI assistant? Ask it to summarize this in ChatGPT, Claude, or Perplexity.