Wan 3.0 vs Alibaba Wan 2.7: A 33% Premium and the End of Open Weights

Doubling the maximum clip length is the headline of Wan 3.0, and it is the least interesting number in this comparison. The one that decides the upgrade is 33%: at 1080p, Alibaba Wan 2.7 is listed at $9.00 per minute of generated video and Wan 3.0 at $12.00, so every finished minute of 1080p work costs three dollars more on the newer model. What that buys is real — a single continuous shot of up to 30 seconds instead of 15, and a document input no other model in the family accepts. What it costs is less obvious, and one part of it is not on any launch slide: the open-weights chapter of the Wan line is closed.

A note on where the numbers come from, because this matchup is unusually easy to get wrong. Alibaba's per-second rates are vendor list prices, published on Alibaba Cloud and repeated in its own comparison post. The quality figures below are not Alibaba's — they come from Artificial Analysis's Video Arena, an independent blind-comparison leaderboard that now carries both generations side by side. Alibaba has published no benchmark table for either model. Where a claim is the vendor's own, it is labelled; where it is the arena's, it carries a capture date, because Elo on that board drifts as votes accumulate.

The upgrade arithmetic, at every resolution

Both models bill per second of generated video, which makes the comparison unusually clean. Alibaba's list rates for Wan 3.0 are $0.05/s at 480p, $0.10/s at 720p and $0.20/s at 1080p. Wan 2.7 sits at $9.00 per minute at 1080p on the independent pricing column Artificial Analysis tracks — $0.15/s — with the same $9.00/min rate on the board's older snapshot.

Run the clip lengths:

• A 15-second 1080p clip — Wan 2.7's ceiling, half of Wan 3.0's — costs about $2.25 on Wan 2.7 and about $3.00 on Wan 3.0. Same length, same resolution, a third more.

• A 30-second 1080p clip — only Wan 3.0 can make it in one pass — costs $6.00 there. On Wan 2.7 the equivalent is two 15-second generations plus an edit, so $4.50 of raw generation plus the assembly time and the continuity work at the join.

• A 30-second 480p draft costs $1.50 on Wan 3.0, which is where the new model is genuinely cheap: a full half-minute concept pass for less than a single 15-second 1080p render on either model.

That third line is the one teams underrate. The 480p tier does not exist on Wan 2.7 at all, and a cheap draft tier changes how many candidates you can afford to throw away — 30 seconds of 480p at $1.50 is a price point where "generate four, keep one" stops being an extravagance.

One more pricing detail worth catching while it lasts: Alibaba attached a limited-time 30% launch discount on Alibaba Cloud's Bailian and Qwen platforms running through 23 September. On a discounted 30-second 1080p clip that takes $6.00 to roughly $4.20. If you are evaluating right now, the effective gap between the two generations is narrower than the list rates suggest, and it widens again on the 24th.

The billing asymmetry almost nobody mentions

Wan 2.7 is not one model with one price. It shipped on 3 April 2026 as a four-model suite — text-to-video, image-to-video, reference-to-video and video-edit — and two of those four bill differently from the other two.

• wan2.7-t2v and wan2.7-i2v — billed on output duration only. Region matters: the international Singapore rates are $0.10/s at 720p and $0.15/s at 1080p, while the Beijing and Tokyo endpoints list $0.086/s and $0.143/s for the same work.

• wan2.7-r2v (reference-to-video) — billed on the input video duration, capped at five seconds, plus the output duration. Feeding a five-second reference into a fifteen-second generation bills for twenty.

• wan2.7-videoedit — billed on output only. The input footage is free.

That single distinction flips the cost comparison for editing workloads. If your work is instruction-based editing of existing footage rather than generation from scratch, Wan 2.7's editing variant charges nothing for the material you feed it, and the effective price per finished clip can land well below the headline 1080p rate. Wan 3.0's all-in-one task routing does not expose the same lever, because it prices the generation rather than the task. Check which variant your quote is actually for before comparing per-second numbers — a flat per-second rate hides this entirely.

Timing also matters more than usual for Wan 2.7. Alibaba Cloud has published a Model Studio retirement notice (ID 118434) taking several older models offline on 10 October 2026, and the list includes wan2.7-r2v and wan2.7-image. This is variant-level, not a shutdown of the 2.7 generation — wan2.7-t2v and wan2.7-i2v are still listed with live rates and pages updated as recently as 11 September — but if reference-to-video is the reason you were considering 2.7, that route has a date on it. No comparable notice exists for Wan 3.0.

Where the arena puts them

Wan 2.7 went six months without a public score. The current Artificial Analysis text-to-video leaderboard (with-audio board, captured 18 September 2026) lists both generations, and the ordering is unambiguous:

• Wan 3.0 — #2 overall, Elo 1,229 (95% CI ±9), 6,011 comparisons, $12.00/min.

• Wan2.7-260612 (the June build) — #7, Elo 1,149 (±6), 16,694 comparisons, $9.00/min.

• Wan 2.7 (the April build) — #10, Elo 1,098 (±7), 4,856 comparisons, $9.00/min.

• Wan 2.6 — #27, Elo 1,030, $9.00/min.

The confidence intervals are worth reading before the ranks, because they decide which gaps are real. Wan 3.0's ±9 band and Wan2.7-260612's ±6 do not overlap: the 80-point lead of the newer model over the best 2.7 build is a resolved difference, not noise. The 51-point gap between the two 2.7 builds is also resolved. What is not resolved is the margin at the top of the board — Wan 3.0 sits three points behind the leader with a ±9 band, which is a tie in everything but presentation.

The Wan 2.7 duplication is not an error, and it is the most useful detail on the board. Alibaba shipped more than one build under the 2.7 name, and the later June revision sits 51 Elo above the earlier snapshot at the same price. Two lessons follow. First, "Wan 2.7" as a procurement line item is ambiguous — ask which build. Second, the gap between the best 2.7 and Wan 3.0 is 80 Elo, not the 131 the headline comparison suggests. That is still a decisive margin on a board where the whole 36-model spread runs about 336 points, but it is a smaller upgrade than the marketing implies.

The picture changes by task. On Artificial Analysis's video-editing leaderboard — a different arena with a different prompt set — Wan 3.0 is #1 at Elo 1,190 with 4,647 votes, and Wan 2.7 is #5 at 1,080. Instruction-based editing is the one capability the two generations genuinely share on paper, since Alibaba carried it forward rather than reinventing it, and the newer model still wins that head-to-head by 110 points. If editing is your workload, the upgrade case is stronger than the text-to-video numbers alone suggest.

The 30-second question nobody asks before buying

Thirty seconds is worth paying for only if you can use thirty seconds. Two failure modes are worth naming.

The first is that a longer clip is more expensive to discard. A four-attempt search for one good 5-second insert costs $4.00 at 1080p. The same four attempts at 30 seconds costs $24.00. Longer generations fail in more ways — drift accumulates, the beat around 20 seconds sags, a face changes between the two halves of the shot — so the discard rate on long clips is not lower than on short ones. Budget for iteration, not for one hero render.

The second is that the extra time only helps when continuity helps. A 30-second commercial cut into six deliberate 5-second shots gains nothing from a longer ceiling and pays 33% more per second for it. A one-take product demonstration, a continuous camera move, a monologue that has to land in a single shot — those are the workloads where the doubling is the whole point, and where assembling two 15-second clips on Wan 2.7 starts to show at the join: a light shift, a motion reset, a pacing restart.

My read: if more than about a third of your output is intentionally cut into short pieces, Wan 2.7 is still the better buy at $9.00/min. If your briefs regularly call for continuous takes, the extra $3.00 a minute is cheaper than the edit that fixes a bad join.

What else actually changed, and what did not

Alibaba's own comparison post is the cleanest statement of the delta, and it is narrower than the coverage suggests.

• Duration — Wan 2.7: 2 to 15 seconds per generation. Wan 3.0: up to 30 seconds in one pass, plus smart duration recommendation and a video-extension tool for chaining past the ceiling.

• Inputs — Wan 2.7: text, image, video, audio. Wan 3.0: the same four plus documents and web pages — DOC, DOCX, XLS, XLSX, PPT, PPTX, PDF, TXT, MD and public URLs. Document-to-video is genuinely new to the family and has no equivalent on the older model.

• Reference limits — Wan 2.7: up to 5 subject images and 5 reference clips. Wan 3.0: up to 10 reference images, 5 reference videos and 5 reference audio clips, within an overall ceiling of 20 reference assets per request.

• Editing — instruction-based video editing was introduced in Wan 2.7 and carried into Wan 3.0 rather than replaced. This is the one capability where the older model is not simply behind.

• Consistency — Alibaba claims finer control over facial detail, hair, clothing, accessories, props and spatial relationships on Wan 3.0, holding across longer clips. That is an unevaluated vendor claim; the arena scores above are the closest thing to a check on it.

• Audio — native dialogue, music and sound effects in the same pass on both, on by default in Wan 3.0 and disableable with a flag. Alibaba's own announcement names audio quality and on-screen text accuracy as areas still needing work — the vendor's own caveat, not a reviewer's.

• 4K — do not buy this on resolution. Alibaba's materials commit to 480p, 720p and 1080p. Native 4K appears on third-party marketing pages, not in Alibaba's announcement.

Where Wan 2.7 still wins, and where both break

The comparison pages treat this as a one-way upgrade. It is not, on three counts.

The first is a vendor-adjacent claim that only exists for the older model. Wan 2.7 was announced as #1 on DesignArena's video-to-video editing board at Elo 1,334, ahead of the next entry by 68 points — a figure published in April 2026 and based on crowdsourced votes rather than a controlled lab evaluation. Wan 3.0's instruction-editing capability is carried forward rather than reinvented, so there is no equivalent published #1 claim for it. Treat the DesignArena number as a marketing-adjacent result, not a lab measurement, but note that nobody has published a competing claim for the newer model either.

The second is that the failures are different in kind, and the older model's are better documented. Reviewers of Wan 2.7 report functional but not frame-perfect lip sync: it drifts above roughly 150 words per minute, collapses multiple simultaneous speakers into a single voice, and lets music overwhelm dialogue unless you duck it manually in post. Generation-side artifacts cluster around heavy motion blur on chrome and grille detail, and detail loss on head turns under strong facial shadow. Identity drift over long or complex scenes is the most consistently reported problem, with a good first frame masking a face that changes later.

Wan 3.0 addresses the duration half of that story — a 30-second single pass has fewer joins for identity to drift across — but Alibaba's own launch materials name audio quality and on-screen text accuracy as areas still needing work. Longer clips also accumulate drift within the shot rather than at the cut. Neither generation is a finished audio tool, and both vendors say so about themselves.

The third is resolution. Wan 2.7 tops out at 1080p, and so does Wan 3.0. One widely syndicated comparison claims Wan 2.7 generates at 4096×2160; Alibaba's own documentation lists only 720p and 1080p as billable tiers, and the Wan 3.0 announcement commits to 480p, 720p and 1080p. If native 4K is a requirement, this upgrade does not solve it and neither generation is the answer.

The thing that changed for anyone who self-hosts

This is the part that does not show up in a spec comparison, and it matters more than any single feature.

Wan 2.1 and Wan 2.2 shipped full weights under Apache 2.0. Wan 2.5 was announced as open and never released weights. Wan 2.6 was a closed API. Wan 2.7 was only partially open. Wan 3.0 continues that direction: it is a metered, closed API, and Alibaba's launch materials say nothing about releasing weights. If you are on Wan 2.2 today and waiting for a newer downloadable checkpoint, you are waiting on something Alibaba has not promised, against a four-release trend pointing the other way. The honest planning assumption is that the last self-hostable Wan flagship is Wan 2.2.

Calling either one, and what the bill should look like

Both models are reachable through Alibaba's own consoles — Model Studio on Alibaba Cloud, and Qwen Cloud — and through third-party creative tools that have integrated them. Wan 3.0's API opened fully at the 24 August launch, lifting the application gate that had been in place since the 6 August beta.

Where the per-second billing style creates a real risk is in the quoting. When a model is priced per second per resolution, the number you are handed by an intermediary is often not the vendor's number, and the markup hides inside a rate that looks like a list price. The check is simple: ask whether the rate you are being quoted is Alibaba's published rate for that resolution, and compare against the list of $0.05 / $0.10 / $0.20.

For the models we do route, that question has a structural answer rather than a promise attached to it. OrcaRouter passes provider list price through at 0% markup across 200+ models on one API key, which means a vendor price cut — the kind currently running on Wan 3.0 through 23 September — is live on our side the same day rather than after a contract cycle. Wan 3.0 is not one of the models we route; it is available through Alibaba's own platforms and through third-party platforms that have integrated it. We would rather say that plainly than imply otherwise, because the pricing lesson is the same either way: the model is not where the money leaks, the billing is.

If you are running the comparison as a build decision rather than a research exercise, the shape of the test that answers it is three prompts — one continuous 25-second take, one six-cut 30-second sequence, and one edit instruction applied to existing footage — run on both generations at 720p, scored on cost per clip you would actually ship. At $0.10 and $0.15 a second those runs cost a few dollars and settle a question that no leaderboard position resolves for your particular footage.

Which one to call

Wan 3.0 is the better model and it is not the better buy for everyone. The arena margin over the June 2.7 build is 80 Elo, the price premium is 33%, and the capability delta is one genuinely new input type, a doubled ceiling, a wider reference budget and a cheap draft tier.

Choose Wan 3.0 if your work involves continuous takes longer than 15 seconds, if your briefs are driven by documents or web pages rather than images, if you need to hold more than five references across a scene, or if you want a 480p draft tier to iterate cheaply before committing to a 1080p render. Choose Alibaba Wan 2.7 — specifically the June build, and ask for it by build — if your output is deliberately cut into short shots, if instruction-based editing is the main workload rather than a supporting one, or if you are running high volumes at 1080p and the three dollars a minute compounds. And if self-hosting is a hard requirement, neither of these is the answer; Wan 2.2 still is.

The open question for the next quarter is whether Alibaba keeps the two lines separate. Wan 2.7 at $9.00/min with instruction editing intact is a perfectly serviceable product and a cheaper floor for the family. Keeping it alive is how Alibaba holds the price-sensitive end of the market while Wan 3.0 takes the top. The thing to watch is not a new model but the 2.7 price: if it moves, the upgrade arithmetic in this article moves with it, and the 30% launch discount expiring on 23 September is the first data point.