A tiered cost strategy for Seedance variants

Updated 2026-10-02

The cheapest Seedance variant is not the cheapest way to make a video, and the most expensive one is not the most expensive way either. What matters is how many full-price generations you pay for. This page is a plan for assigning variants to stages so the full-price model only renders shots you have already approved.

Why variants behave like stages

Seedance is sold as separate model ids that share a request shape: bytedance/seedance-2.0-mini, bytedance/seedance-2.0-fast, bytedance/seedance-2.0 and bytedance/seedance-2.5, plus an older 1.5 Pro row. Swapping between them is a one-string change, so a pipeline can run the same inputs through different variants at different stages. Per-second rates differ across variants and across resolution tiers; the live table has current figures:

Per second, before VideoRouter's 2% platform fee. For tiered models each row compares the resolution tier with the widest host-to-host gap. Built 2026-10-02 from the live catalog.

The three-stage workflow

1. Explore on Mini or Fast. Test prompts, camera moves, start images and reference sets. Use the lowest resolution tier that lets you judge composition and motion, and short durations.
2. Select. Choose the shots worth finishing. Keep the exact prompt, seed-equivalent inputs and reference arrays that produced them.
3. Finalise on 2.0 or 2.5. Re-submit only the approved shots to a full variant at the delivery tier.

The same start_image_url or reference arrays work across the stages, which is the practical benefit of a shared request shape.

The math, symbolically

Let r_m, r_f, r_full be the per-second rates for your chosen tiers on Mini, Fast and a full variant. Let d be the clip length, n the number of shots you deliver, e the number of exploration generations per shot, and k the number of full-price attempts per approved shot.

all_full   = n x e x d x r_full
tiered     = n x e x d x r_m   +   n x k x d x r_full

Tiering wins when:

e x r_full  >  e x r_m  +  k x r_full
=>  e x (r_full - r_m)  >  k x r_full

In words: the saving on each exploration attempt, times the number of attempts, must exceed the cost of the full-price tries you still need. If you rarely explore (small e) or the draft variant is poorly correlated with the final (large k after drafting), the benefit shrinks. Multiply both sides by your platform fee of 2% if you want billed numbers; it does not change the inequality.

You do not need invented numbers to use this. Pull r_m and r_full from the table, count your own e and k from a week of logs, and evaluate the inequality. The per-minute guide extends it with retry multipliers and host spread.

When not to downgrade

• The failure you are testing only appears in the full variant. Fine texture, small text, hands and logos can differ between variants; a draft will not settle them. For brand-critical shots, plan on more full-price attempts (a larger k) or skip the draft stage.
• Few iterations. If you generate one or two clips per shot anyway, the extra draft stage mostly adds latency and engineering.
• Consistency across a sequence. Drafts from one variant and finals from another will differ in look. If shots must match each other, finalise the entire sequence on the same variant rather than mixing.
• Latency-critical user-facing features. Fast exists for turnaround. Use it as the product, not only as a draft, if speed is the requirement.
• Reference-heavy shots. Support for multi-reference inputs varies by host and variant; confirm that the draft model accepts the same arrays as the final before relying on it.

A variant ladder you can reuse

If you want a default, use this: Mini or Fast for exploration and for anything a customer will see only as a thumbnail or preview; the full 2.0 variant for approved shots when the clip will be shown at delivery quality; 2.5 for the final pass when you want the newest model and the host table shows a rate you can afford. Revisit the ladder whenever a host reprices, because the cheapest variant at one tier can overlap the next variant's rate at a different host. Read the live table before each large batch rather than relying on last month's note.

Controls that keep the plan honest

1. Route drafts and finals through separate API keys with separate monthly spend limits. A leak in the draft loop then cannot eat the finals budget.
2. Log requested duration, variant, tier and resulting job id per call. You need e and k from real data.
3. Never re-submit a slow job; polling is free and billing is at creation. Failed upstream jobs are not billed.
4. Decide host policy per stage: float for drafts (cheapest healthy host with failover), pin only if a capability requires it.

The five-lever guide covers the surrounding tactics, and the cost page shows 5s, 10s and 60s clips at the cheapest host. To start, create a key and send the quickstart request to Mini.

Frequently asked questions

Keep reading

• How to Reduce Your Seedance API Cost — Five Practical Levers
• Seedance API Cost Per Minute: Budget Formula & Retry Math
• Seedance vs Kling API Pricing: A Fair Cost Comparison
• Seedance 2.5 Price Per Resolution Tier: How to Read the Table