A question most comparison articles skip: how do AI video generation APIs actually charge, and what does one finished clip really cost?
On August 24, Alibaba's Wan3.0 opened its API on Model Studio (see Wan3.0 Officially Launches), priced at 1.2 RMB/second for 1080P with a single-segment maximum of 30 seconds. That looks like just another pricing announcement - until you put it into the ledger of "total cost of one 30-second 1080P clip." Then the picture flips: the single-segment duration cap is becoming a hidden cost more expensive than the per-second price itself.
Scope first: this site already ran a capability-focused video generation comparison on August 6 (Sora Bows Out: Veo 3.1 vs Kling 3.0 Fight for the Crown). This article does not repeat the "who has the best quality" story - it only does the money: billing models, price provenance, single-segment duration caps, the true total cost of one 30-second clip, and API integration differences. Every price is tagged with its source and collection basis (official pricing vs aggregator quotes); overall this is a representative comparison assembled from public pricing and third-party hands-on data - not our own stress test - so check each vendor's official page before you commit.
1. The Pricing Baseline: How Six Vendors Bill
Here are the billing models of six mainstream, generally available video generation APIs. Note the provenance varies a lot: some are first-party official prices, some are aggregator quotes (which may include resale markup or discounts) - each is tagged in the table.
| Platform / Model | Billing model | Representative unit price (basis and date) | Single-segment cap |
|---|---|---|---|
| Wan3.0 (Alibaba Cloud Model Studio) | Per second | 480P 0.3 RMB/s, 720P 0.6 RMB/s, 1080P 1.2 RMB/s (official pricing, live 2026-08-24; 30% off promo through 09-23) | 2-30 seconds (integers) |
| Kling 3.0 (Kuaishou) | Per second of final generated video | 1080P $0.14/s, 720P $0.112/s, plus a $0.084/s tier; Turbo tier ships native audio (official pricing page, relayed via CSDN 2026-08-16) | Typically 5/10-second tiers |
| Vidu Q3 pro (Shengshu) | Per second | 1080P 1 RMB/s, 720P 0.9375 RMB/s, 540P 0.4375 RMB/s (aggregator now.cn quote, updated 2026-03) | 1-16 seconds |
| Sora 2 (OpenAI) | Per second | Text-to-video 720P 0.71 RMB/s; pro 720P 2.14 RMB/s, pro 1080P 3.56 RMB/s; image-to-video same price (aggregator quote) | Typically 4-12-second tiers |
| Hailuo 2.3 (MiniMax) | Per clip | 6s/768P 2 RMB, 10s/768P 4 RMB, 6s/1080P 3.5 RMB; Hailuo-2.3-Fast cheaper (6s/768P 1.35 RMB, 6s/1080P 2.31 RMB) | About 10 seconds |
| Seedance 2.0 (ByteDance / Volcano Engine) | Per second | 1 RMB/s (the "1 RMB per second" widely reported by media; predecessor 1.5 pro bills per million tokens - 16 RMB with audio / 8 RMB without) | Depends on version |
Three first-glance conclusions:
- Billing granularity splits into three camps: per second (Wan3.0, Kling, Vidu, Sora, Seedance), per clip (Hailuo), and per credit (Runway Gen-4, 1 second = 5 credits; third-party testing put 8 seconds at about $0.4). The coarser the granularity, the more expensive your experiments - Hailuo bills 2 RMB minimum per clip, failed takes included.
- Resolution tiers within one vendor can differ 4x: Wan3.0 spans 0.3 to 1.2 RMB/s from 480P to 1080P; Sora 2 spans 0.71 to 3.56 RMB/s from standard 720P to pro 1080P - 5x. The resolution tier is the first money-saving lever, and the ledger below will use it.
- Previous-gen prices haven't collapsed: Wan2.6 runs about 1 RMB/s for 1080p on aggregators (2026-03 basis), Wan2.5 about 1.095 RMB/s. In other words, Wan3.0's first-party official 1.2 RMB/s is roughly on par with what the previous generation resold for - a new model with no premium, which is rare among video APIs; during the 30%-off window it drops to 0.84 RMB/s.
One more detail worth recording: Wan3.0 prices Beijing, Tokyo, and Frankfurt regions identically, with Singapore slightly higher (480P 0.37471 / 720P 0.74942 / 1080P 1.49884 RMB/s). Teams hedging across multi-region deployments pay about 25% more in Singapore.
2. The Ledger: What One 30-Second 1080P Clip Actually Costs
This is the core of the piece. We pick 30 seconds at 1080P as the benchmark because it is the typical length of a Douyin/WeChat Channels talking-head short or a product demo segment. The formula:
Total cost of one 30-second 1080P clip = 30s ÷ single-segment cap (round up = segment count) × per-segment price + stitching and retry overhead
Each calculation, re-derived line by line (USD converted at 1≈7.1-7.2 RMB, 2026-08 basis; excludes gacha retries and stitching labor - that ledger comes next):
| Platform / Model | Calculation | Total for one 30s 1080P clip | Segments and stitching cost |
|---|---|---|---|
| Wan3.0 | 30s × 1.2 RMB/s | 36 RMB (25.2 RMB during the 30%-off window) | One segment, zero stitching |
| Kling 3.0 | 30s × $0.14/s = $4.2 | About 30 RMB | Needs 10-second tiers × 3 stitched segments |
| Vidu Q3 pro | 30s × 1 RMB/s | About 30 RMB | 16s cap, needs 2 segments (16s + 14s) stitched |
| Sora 2 pro | 30s × 3.56 RMB/s | About 107 RMB | Typically 4-12s per segment, at least 3 stitched; access and payment from mainland China add further cost |
| Hailuo 2.3 | 6s/1080P 3.5 RMB × 5 segments | 17.5 RMB (768P 10s × 3 = 12 RMB; Fast tier 1080P 2.31 × 5 ≈ 11.6 RMB) | 5-6 stitched segments; more segments, more style-drift risk |
| Seedance 2.0 | 30s × 1 RMB/s | Around 30 RMB | Segment count depends on the version tier |
Three overlooked truths hide in this table:
Truth one: the single-segment duration cap is an underestimated hidden cost. Kling, Vidu, Seedance, and Sora all show per-second prices in the same band as Wan3.0 - but for a 30-second film you must cut it into 2-6 separately generated segments and stitch them in an editor. Stitching costs more than labor: visual-style consistency across cuts, camera-move continuity, and character stability are all gacha-grade luck. The more segments, the faster the probability that "every second of the whole clip is usable" decays multiplicatively. Wan3.0 writing 2-30 seconds single-segment output into its hard specs prices this hidden cost at zero - that's its real enterprise selling point, far more concrete than quality benchmarks.
Truth two: iterate on the cheap tier, render on the expensive tier - the biggest saving lever. Take Wan3.0: 30 seconds at 480P is only 9 RMB (6.3 RMB discounted), at 720P 18 RMB (12.6 RMB discounted). The battle-tested playbook: use 480P/720P to lock down prompts, storyboard, and input assets, and only render the final cut at 1080P. Ten rounds of prompt tuning cost 360 RMB if run entirely at 1080P, but 90 + 36 = 126 RMB with the two-tier approach - 65% saved. The same playbook works on every platform that bills per second and tiers by resolution.
Truth three: the spread can hit 20x. Third-party hands-on testing (2026-07-14, via Baijiahao) ran the same 8-second 1080P prompt across platforms: Runway Gen-4 about 2.9 RMB (credit system), Pika about 2.5 RMB per clip, while Sora 2 pro costs 3.56 × 8 ≈ 28.5 RMB for the same spec. The same 8 seconds at 1080P spans roughly 10-20x between platforms - but cheap isn't the same as good value, because retry costs from unusable takes claw part of the gap back. That's the next section.
3. Putting the Gacha Back In: Real Cost Needs a Multiplier
The same third-party test delivered a brutal number: even Kling- and Runway-class models have roughly a 15-20% chance per generation of quality below bar. In expectation, you pay a 15-20% "scrap tax" on every finished clip.
The more aggressive, institutionalized version is to make the gacha a pipeline: fire 3 generations per task in parallel and auto-pick the best with a scoring model (VLM-based). Measured result: satisfaction rose from 65% to 92% - at the price of tripling the API bill. Applied to the 30-second benchmark:
| Strategy | Wan3.0 (1080P) | Kling 3.0 (1080P) |
|---|---|---|
| Single shot, no filtering (expected 1.2x retries) | 36 × 1.2 ≈ 43 RMB | 30 × 1.2 ≈ 36 RMB |
| 3 parallel takes + auto-pick | 36 × 3 = 108 RMB | 30 × 3 ≈ 90 RMB |
See that? "Generate 3 at once" instantly erases Wan3.0's price advantage - which is exactly when you should mix tiers instead: run the early rounds of parallel gacha at 480P (3 × 9 = 27 RMB), then re-run the winning take's exact parameters once at 1080P, for a total of 27 + 36 = 63 RMB - about 40% cheaper than gacha-ing 3 takes directly at 1080P. Caveat: "re-run with the same parameters" depends on the platform reproducing prompts and inputs stably; the gacha selects composition and motion, not pixel-level identity.
4. Choose by Scenario: Five Routes
- Need a 15-30 second finished clip in one piece (talking-head short, product demo, PPT-to-video): Wan3.0 is currently the only single-segment 30-second option - 36 RMB (25.2 RMB discounted) buys zero stitching, plus direct PDF/PPT/DOCX file input. The full walkthrough is in our Wan3.0 Hands-On SOP.
- Need one high-quality 5-10 second shot (ad creative, title sequence): Kling 3.0 or Seedance 2.0. Kling's Turbo tier ships native audio; Seedance at 1 RMB/s generates fast (3-8 minutes), and Volcano Engine's API is near-OpenAI format with the lowest integration cost - third-party testing found it 20-40% cheaper than Kling.
- Extremely budget-sensitive, 768P acceptable: Hailuo 2.3 - 4 RMB per 10s/768P clip, 30 seconds in 3 segments for 12 RMB total, the cheapest finished-clip plan in the table; the Fast tier's 1080P runs about 11.6 RMB across five segments. The price is 5-6 stitched segments and the consistency risk that entails.
- Already on the OpenAI stack and want top-tier look: Sora 2 pro 1080P at 3.56 RMB/s breaks 100 RMB for 30 seconds. It is the table's only luxury item, with mainland access and payment rails as extra time and compliance cost.
- High-volume short-video matrix (hundreds of clips a day): don't stare at unit price - run three ledgers first: segment count imposed by the single-segment cap × gacha multiplier × stitching labor. At scale, labor quickly outgrows the API bill, and "single-segment output" should be weighted first.
5. Three Traps to Avoid
- API formats aren't unified - three vendors, three integrations. Kling uses its own format (not OpenAI-compatible) and needs a dedicated adapter layer; Volcano Engine is closest to OpenAI's format and can be wired up in half a day; Runway is REST but bills in credits, so you reconcile two units of account. Teams mixing platforms should wrap a unified "submit-poll-fetch" abstraction in their own service instead of coupling business code to any vendor's SDK. Also note Wan3.0 currently has no batch inference, no Function Calling, no context caching, and
fileandlinkinputs can't be used together - if you hoped to batch-run an entire asset library in one API call, think again. - Async jobs need real polling and timeout logic. Mainstream platforms take 3-15 minutes, all async submit-plus-poll. Practical settings: poll every 10-15 seconds, cap at 20 minutes, and log-plus-reconcile timed-out jobs - billing follows "final generated seconds" (Kling states this explicitly), and "task failed" versus "task succeeded but you don't want it" are two different refund paths; don't conflate them.
- Gacha retries are the number-one cost doubler. A 15-20% below-bar rate means an unattended auto-retry loop quietly inflates the bill, while "3 takes plus auto-pick" is an explicit ×3. The right posture is the two-tier playbook from section 3 - cheap-resolution parallel gacha, expensive-resolution single final render - plus cost accounting per finished clip (scrap included). Run it for two weeks and you'll know your business's real "cost per second" versus the sticker price.
One-line closer: unit price is the visible card; segment count, gacha, and stitching are the hidden ledger - price all three in, and you get the real price of API video generation.
FAQ
Q1: Wan3.0 at 1.2 RMB/s versus Kling at $0.14/s - which is actually cheaper? A1: After conversion they're nearly identical ($0.14 ≈ 1 RMB/s); 30 seconds lands around 30 RMB either way. The difference is segment count: Wan3.0 outputs 30 seconds in a single segment - one task, 36 RMB (25.2 RMB discounted); Kling needs the 10-second tier split into 3 segments - three tasks totaling about 30 RMB, plus stitching labor and style-consistency risk. In production-scale B2B use, the hidden value of "one task" usually beats a few cents per second.
Q2: Why does Sora 2 work out to 100+ RMB while others are so much cheaper? A2: Because only Sora 2 pro has a 1080P tier, and at 3.56 RMB/s it is the priciest in the table - about 107 RMB for 30 seconds; segments typically cap at 4-12 seconds, so longer films must be stitched. Its quotes also come from aggregator platforms, and mainland access/payment adds further channel cost. It sells look ceiling, not value.
Q3: Aggregator quotes are lower than official prices - can I just use them? A3: They save effort, but check three things: quote collection date (the now.cn quotes cited here are from 2026-03, refreshed monthly), whether resale markup or discounts are baked in, and who you call when it breaks. Every aggregator quote in this article is tagged with its basis; for production, use official APIs as primary and aggregators as hedged backups.
Q4: Is there an actual money-saving combo? A4: Yes - the two-tier play: tune prompts, storyboards, and input assets entirely on cheap 480P/720P (Wan3.0's 30 seconds at 480P is just 9 RMB), then render the final once at 1080P. Ten iterations cost 360 RMB all-1080P versus 126 RMB two-tier. Same for gacha: run 3 parallel takes at low resolution, then re-run the winner's exact parameters at high resolution - roughly 40% cheaper than gacha-ing 3 takes at 1080P.
Q5: Will these prices change? When is the best time to get in? A5: They will. Video API prices are dropping fast (from Wan2.5's 1.095 RMB/s to Wan3.0's official discounted 0.84 RMB/s is a 23% cut in six months). The near-term window is explicit: Wan3.0's 30% off on Model Studio and Qwen AI platforms runs through 2026-09-23 - if you have a hard need for 30-second single-segment clips, this month is the cheapest time to experiment. Longer term, build budget elasticity around "cost per second drops ~20% every six months," and never hard-code today's prices into an annual contract.
References
- Alibaba Cloud Model Studio Wan3.0 official pricing and docs (live 2026-08-24, 30% off through 09-23): https://developer.aliyun.com/article/1757799 , https://help.aliyun.com/zh/model-studio/text-to-video-guide/
- Sina Finance / Sanyan Technology: Wan3.0 launch coverage (Singapore-region pricing, 30-second single segment): https://baijiahao.baidu.com/s?id=1874379840707406998
- Kling 3.0 API official pricing page (relayed via CSDN, 2026-08-16): https://blog.csdn.net/qq_40374604/article/details/163803625
- Vidu billing documentation (aggregator now.cn, 2026-03 basis): https://help.now.cn/aimodel/billing/20260320153307/
- Third-party hands-on test "How to Choose an AI Video Generation API" (2026-07-14; source of the 20x spread, 15-20% gacha rate, three traps, and Runway/Pika data): https://baijiahao.baidu.com/s?id=1870653220898565959
This is a representative comparison assembled from public pricing and third-party hands-on data (prices collected as of 2026-08-25; USD converted at 1≈7.1-7.2 RMB), not our own stress test; prices and duration caps may change at any time - verify against official pages before purchasing.