Wan 3.0 vs Wan 2.7: Twice the Length, Half the Price

Wan 3.0 doubles the maximum clip length to 30 seconds and costs half as much per second as Wan 2.7. A generation upgrade that cuts the rate rather than raising it is rare enough to be worth checking twice, so: on Ofox, alibaba/wan-3.0 bills $0.05 per output second and alibaba/wan-2.7 bills $0.10, read from the live catalog on 4 September 2026.

TL;DR

• Half the price. $0.05 per second against $0.10. A 5-second 1080p clip is $0.25 instead of $0.50.
• Twice the ceiling. 30-second maximum against 15. Both still start at 2 seconds.
• One model instead of several. Wan 2.7 splits tasks across model strings on Alibaba Cloud; Wan 3.0 infers the mode from your inputs.
• More output shapes. 480p added below 720p and 1080p, and three more aspect ratios plus an adaptive default.
• The migration is a string change. Same Ofox endpoint, same parameter names.

Wan 3.0 vs Wan 2.7: What Actually Changed

Prices and attributes are the pricing and video_attributes fields on the live Ofox /v1/models response; release dates are from the dated aliases wan-2.7-20260414 and wan-3.0-20260824. Read 4 September 2026.

Wan 2.6 sits alongside 2.7 at the same $0.10 rate and the same 15-second ceiling, with one fewer mode. If you are on 2.6, everything below applies to you at least as strongly.

The Price Cut Is Real, and It Is the Point

Half the rate on every clip, forever, not an introductory window. Alibaba runs a separate 30% promotional discount on its own platform through 23 September 2026, but the Ofox flat rate is not that discount and does not expire with it. The Wan 3.0 pricing guide has the route-by-route breakdown, including why Alibaba’s own bill varies by resolution while this one does not.

There is no caching, no batch tier and no variable output length to model here. The rate times the seconds is the bill, so the halving passes through cleanly at any volume.

The 15-Second Wall Is Gone

Wan 2.7 stops at 15 seconds. Wan 3.0 goes to 30. For anything longer than 15 seconds on 2.7, the workaround was to generate in segments and stitch, which costs you continuity at every seam and engineering time around the joins.

Wan 3.0 also handles continuation directly, and the rule there is worth reading before you plan around it: when you pass input video, the input duration plus the output duration together cannot exceed 30 seconds, per Alibaba’s Wan 3.0 API reference. So 30 seconds is a total budget, not a per-call output allowance on top of whatever you feed it.

One Model Instead of a Model per Task

This is the structural change, and it is easy to miss because it does not appear in a price table. On Alibaba Cloud, Wan 2.7 exposes tasks through separate model strings and reference pages: wan2.7-t2v for text-to-video, with first-frame, first-and-last-frame and continuation documented separately. Wan 3.0 collapses that into wan3.0-video, taking text, images, video and audio mixed in one request, and covering text-to-video, image-to-video, reference-to-video, editing and continuation.

Through Ofox the difference is smaller than it sounds, because both models already sit on /v1/videos with the same parameter set and the mode follows from your inputs: frame_images selects first-frame or first-and-last-frame, input_references selects reference-to-video. What changes is that on 3.0 the same behaviour is native to a single upstream model rather than a gateway smoothing over several.

The Migration Is One String

- "model": "alibaba/wan-2.7"
+ "model": "alibaba/wan-3.0"

Nothing else in the request has to change. Both models accept prompt, duration, resolution, aspect_ratio, size, generate_audio, seed, frame_images, input_references, callback_url and provider, and both default to 1080p and to audio on.

Three things to re-check after the swap:

• Aspect ratio. Wan 2.7 supports 16:9, 9:16 and 1:1, and rejects anything else with aspect_ratio "4:3" not supported; allowed: [16:9 9:16 1:1]. Wan 3.0 supports those plus 4:3, 3:4 and adaptive, and adaptive is the default. Anything that relied on an unset aspect ratio behaving a particular way should now pin it explicitly. The Wan 3.0 error reference has the full rejection set for both models.
• Duration assumptions. Code that clamps user input to 15 seconds because 2.7 required it is now leaving half the range unused. Code that assumed a 15-second maximum for cost estimation will under-estimate a 30-second clip by exactly 2x.
• Cost alerts. Any budget alarm calibrated on $0.10 per second will now fire at half the expected spend, or not at all. That is a pleasant direction to be wrong in, but it still means the alarm is not measuring what you think.

When to Stay on Wan 2.7

Honestly, not often, and the usual reason to stay put does not apply here. There is no price penalty for moving, so the only real cases are:

• Prompts tuned tightly against 2.7 output. Motion behaviour and prompt adherence changed between generations, and Alibaba positions 3.0 as a different model rather than a tuned 2.7. If your prompt library is large and validated, re-validation is the actual cost of the upgrade, not the API bill.
• A pinned dated alias for reproducibility. wan-2.7-20260414 gives you a fixed target. If you have an approval process that requires output to be reproducible against a known model version, changing it is a process question, not a technical one.

Neither of those is a price argument, which is the unusual part of this upgrade.

Where This Sits Against the Competition

At $0.05 per second, Wan 3.0 undercuts more than its own predecessor. Seedance 2.5 bills $0.11 per second, which makes it 2.2x the rate, though it accepts substantially more reference material and adds a 21:9 aspect ratio that Wan does not list. If you are picking between vendors rather than between Wan generations, that comparison is the relevant one.

Sources

• https://ofox.ai/models/alibaba/wan-3.0
• https://ofox.ai/models/alibaba/wan-2.7
• https://help.aliyun.com/zh/model-studio/wan3-video-generation-api-reference
• https://help.aliyun.com/zh/model-studio/text-to-video-api-reference

Rates, duration bounds, resolutions and aspect ratios were read from the live Ofox /v1/models endpoint on 4 September 2026 and cross-checked against Alibaba’s Wan 3.0 and Wan 2.7 API references the same day.