Wan 3.0 and Seedance 2.5 are the two models fighting for the top of the 2026 video leaderboard — and the two we get asked to compare most. Here is the full picture: benchmarks, inputs, pricing and which workload each one wins.
The short answer: Wan 3.0 holds the #1 text-to-video spot on Artificial Analysis (with and without audio) and is the cheaper per-second buy on Viddra, starting at $0.25/sec at 480p. Seedance 2.5 is the multi-reference champion — dozens of images, clips and audio files per generation, 30-second takes with synchronized audio, plus extension rounds and partial re-rendering. Wan wins on price and controlled start/end states; Seedance wins on reference-heavy storytelling.
Both run behind the same Viddra key with identical request shapes — change the model field, keep the pipeline. Try them live on the Wan 3.0 and Seedance 2.5 tool pages, or see how they stack against every other model on the models directory.
leaderboardBenchmarks & Leaderboards
On the Artificial Analysis public arena — thousands of blind pairwise votes on real prompts — the October 2026 board has Wan 3.0 first among models with audio (ELO ~1157), with Seedance 2.5 in third at roughly 15 points behind. Wan also takes the top marks on the separately-measured lighting/materials and camera-control criteria, while Seedance 2.5 posts the single strongest image-to-video results in its class.
Two honest caveats: arena positions move weekly, and a ~15-point ELO gap is small enough that prompt fit, latency and price decide most real workloads. Treat the leaderboard as a tiebreaker, not a verdict — then price the actual job, which is what the rest of this page does.
tuneCapability Comparison
| Wan 3.0 | Seedance 2.5 | |
|---|---|---|
| Max single take | 30 seconds | 30 seconds (+ extension rounds) |
| Native audio | Yes, synchronized | Yes, synchronized |
| References per generation | Up to 20 | Dozens — images, clips, audio |
| Frame control | First-frame, last-frame, reference media | Partial re-render, extension rounds |
| Resolutions on Viddra | 480p / 720p / 1080p | 480p / 720p / 1080p (per-token) |
| Billing | Per second | Per token |
| Upstream on Viddra | Alibaba Bailian | fal.ai |
| Model ID | wan3.0-video | seedance2.5 |
GET https://api.viddra.com/v1/modelsView live pricing arrow_outwardThe structural difference is inputs. Wan 3.0's frame control (first/last/reference) makes it the precision tool: you decide where a shot starts and ends. Seedance 2.5's reference depth makes it the consistency tool: feed it a character sheet and a mood-board and it holds identity across a 30-second narrative. Those are different jobs — most teams end up wanting both.
paymentsPricing Breakdown
Wan 3.0 bills per second, which makes it the easiest to cost in advance. Seedance 2.5 bills per generated token (resolution × seconds), so a 720p take costs more than a 480p take of the same length — the pricing page carries the live per-token table.
| Workload | Volume | Wan 3.0 (480p) | Wan 3.0 (720p) |
|---|---|---|---|
| Ad iteration batches | 200 × 6s | $300.00 | $600.00 |
| Product demos | 50 × 15s | $187.50 | $375.00 |
| Long-form hero clips | 20 × 30s | $150.00 | $300.00 |
For volume work, also compare against Viddra's budget tiers — Seedance 2.0 Mini at $0.08/sec and Hailuo H3 at $0.12/sec. A common cost pattern: draft unlimited on the budget tier, re-render only the winners on Wan 3.0 or Seedance 2.5.
check_circleChoose Wan 3.0 When…
— Cost predictability matters: flat per-second billing, cheapest 480p flagship rate on Viddra.
— You need controlled shots: first-frame, last-frame and reference-media inputs for start/end states.
— Long single takes are the job: 30 seconds in one pass, ranked #1 for camera control.
Full integration math in the Wan 3.0 pricing guide, or generate without code on the Wan 3.0 tool page.
movieChoose Seedance 2.5 When…
— Identity consistency across shots is the job: character sheets, product reference sets, mood-boards.
— You want extension rounds and partial re-rendering instead of full re-rolls.
— Narrative multi-shot clips with synced audio in a single pass are the deliverable.
Try it on the Seedance 2.5 tool page, or read how the rest of the field compares in the 2026 API comparison.
helpFrequently Asked Questions
Which model ranks higher, Wan 3.0 or Seedance 2.5?
On the Artificial Analysis public arena (October 2026), Wan 3.0 leads the text-to-video board with and without audio, while Seedance 2.5 sits within about 15 ELO points in third. Both render up to 30 seconds with synchronized audio; the ranking gap is small enough that workload fit matters more than leaderboard order.
How do their input options differ?
Seedance 2.5 is the multi-reference champion: it accepts dozens of reference images, video clips and audio files per generation, plus extension rounds and partial re-rendering. Wan 3.0 takes up to 20 references and adds first-frame, last-frame and reference-media image-to-video on Viddra, which is the stronger fit for controlled start/end states.
What does each cost per second on Viddra?
Wan 3.0 bills per second: $0.25/s at 480p, $0.50/s at 720p and $1.00/s at 1080p. Seedance 2.5 bills per generated token (480p to 1080p) with 30-second audio clips; the live per-token math is on the Viddra pricing page. Both are pay-as-you-go with no subscription.
Can I call both models with one integration?
Yes. Both models run behind the same Viddra API key with the same OpenAI-style request shape — changing the model field switches the renderer. A common production pattern is drafting on Seedance 2.5's multi-reference mode and re-rendering winners on Wan 3.0 for the cheapest 480p takes, or the reverse for reference-heavy storytelling.
Viddra Engineering
More postsThe Viddra engineering team builds the unified API layer that makes frontier video, image and speech models boringly reliable to call.
