Wan 3.0 vs Seedance 2.5: the keep rate where Wan is cheaper

A 5 second 720p clip costs $0.63 on Wan 3.0 and $2.89 on Seedance 2.5 on Sume, so Wan is cheaper per kept clip as long as its keep rate is at least 21.8% of Seedance's keep rate (0.63 divided by 2.89). If you keep half of Seedance takes, Wan only needs to keep about 11% of its takes to cost the same per keeper.

This post is arithmetic about a ratio, not a claim about quality. Neither keep rate is measured here; the numbers in the table are what-if inputs. Prices are from Sume's price tables on main and the Video Generation docs, read 2026-10-10.

The formula

Cost per kept clip equals price per take divided by keep rate. Setting Wan equal to Seedance gives: wan_price / wan_keep = seedance_price / seedance_keep, so wan_keep = seedance_keep x (wan_price / seedance_price) = seedance_keep x 0.218.

Billing is in micros, so the exact 720p amounts ($0.625 and $2.889) give 21.6%; this post uses the rounded cents shown in the price tables, which gives 21.8%. Failed jobs are a separate matter. A job that fails or is canceled is refunded, so the keep rate here means finished takes you reject, which are captured and paid for.

The break-even table

Each row shows a Seedance 2.5 keep rate and the lowest Wan 3.0 keep rate at which Wan costs the same per keeper, for 5 second 720p takes.

What moves the ratio

The 21.8% figure belongs to this pair at this size. At 480p, Wan is $0.32 per 5 seconds and Seedance 2.5 is $1.35, a ratio of 23.7%. At 1080p, Wan is $1.25 against Seedance's $7.11, a ratio of 17.6%. Higher resolution widens the gap because Seedance bills video tokens, which grow with pixels, while Wan's per-second rate grows more slowly.

Length matters less than you would think. Both rows scale linearly with seconds, so the ratio is nearly constant from 5 to 30 seconds apart from cent rounding. Adding a reference video to Seedance changes it: the reference version costs $6.94 for 5 seconds, which pushes the ratio down to 9.1%.

• 480p: Wan needs 23.7% of Seedance's keep rate.
• 720p: 21.8%.
• 1080p: 17.6%.
• Seedance with a reference video at 720p: 9.1%.

Using it in practice

Pick a cheap, honest test: generate the same brief a handful of times on each model and count the takes someone would actually ship. Then compare your own two keep rates to the ratio for your resolution. If Wan keeps more than about a fifth as often as Seedance, Wan wins on cost; if not, the dearer model is the cheaper way to get a usable clip.

Keep in mind that the two models differ in features and look, so this test measures taste as well as cost. A tie on price is not a tie on result, and the break-even only tells you how much worse the cheaper model is allowed to be before it stops saving money.

Two cautions keep the sheet honest. First, count only finished takes in the keep rate; refunded failures are free and should not dilute it. Second, count your own review time if it is significant: a model with a low keep rate makes a person watch more rejected clips, and that cost sits outside the wallet. If reviewing is the expensive part, a smaller ratio of cheap takes can still lose to a dearer model that is right more often. Use the table as a floor, not a recommendation.

A good way to apply it is to decide the Seedance keep rate you expect first, since that is the model you would otherwise pick, then read across to the Wan rate you would need. If that number looks easy to beat, run Wan first and only escalate to Seedance for the shots that do not land. That is a draft-then-final plan with the ratio as its trigger.