How Much Does an AI Short Drama Cost? Per Minute, Per Episode, Per Season (2026)

SinCoSphere EditorialUpdated 2026-09-025 min readCost & pricing

At API list prices a 10-episode × 90-second vertical series costs about $70–260 in model calls if every clip succeeds first time at 720p/768p, and $120–450 at 1080p/2K. Real production multiplies that by 2–3 for regenerations, and a team adds labor that dwarfs the model bill: invideo's own case study reports $1,000 per episode. Video is over 90% of the model bill; parsing, reference art and storyboards together stay under $15.

Why the published numbers disagree by 100×

Three price bands circulate. At the bottom, single-pass model cost: MiniMax H3 bills $0.08 per second at 768p (MiniMax pricing, checked 2026-09-02), so a 90-second episode is $7.20 and a minute under $5. In the middle, creators who count regenerations and a 1080p final land at $30–100 per minute. At the top, invideo's first-party case study puts a 10-episode micro-drama at $1,000 per episode with three people over three days (invideo, vendor claim). Chinese trade press quotes about ¥1,000 per minute (roughly $140 at 7.1 CNY/USD) for polished anime-style AI dramas (Beijing News, 2026-01).

None of these is wrong. They differ on three variables: regeneration rate (first-pass usable clips run 50–70% in our own production data, and every retry is billed again), resolution tier (1080p/2K costs 1.6–2.4× the 720p tier on the same model), and labor (storyboard edits, version picking, cutting). Ask for all three before accepting a per-minute figure.

Bottom-up: a 10 × 90 s season at list prices

Assumptions: 10 episodes, 90 seconds each, 8 segments per episode (one segment = one model call of 10–12 s), so 900 finished seconds and 80 video calls. Every call succeeds once. Prices are official pay-as-you-go list prices unless marked.

Model (channel)	Per second	900 s season	Source, date
MiniMax H3 768p / 2K (official API)	$0.08 / $0.13	$72 / $117	MiniMax, 2026-09-02
Wan 3.0 720p / 1080p (Alibaba Model Studio, international)	$0.10 / $0.20	$90 / $180	Alibaba Cloud, 2026-09-02
Seedance 2.0 720p / 1080p (Volcengine, RMB list, converted at 7.1)	≈ $0.13 / $0.32	≈ $120 / $284	Sina Finance, 2026-03-04; token formula second-hand
Kling 3.0 (two resellers)	$0.084–0.126 vs $0.153	$76–138	Atlas Cloud / BuildMVPFast, 2026-07
Veo 3.1 (aggregator, tier unspecified)	$0.09	$81	BuildMVPFast, 2026-07
Runway Gen-4.5 (subscription, 60 credits per 5 s)	$0.12–0.29	$108–259	Runway, 2026-09-02

Add the non-video stages: one parsing call ($0.10–3.50 depending on the LLM), 30–45 reference images at $0.03–0.13 each, and 10 storyboard calls at $0.05–0.70 each — $2–15 in total, per our cost breakdown. Voice from a preset library is effectively free; compositing is cents. Video is 92–97% of the model bill.

The Seedance row is computed: 1080×1920 × 24 fps ÷ 1024 = 48,600 tokens per second at ¥46 per million (¥28 with video input) ≈ ¥2.24/s. The formula is from a third-party API doc (APIYI, second-hand, self-reported accuracy within 0.1%); Volcengine's own article only describes tiered token pricing. Other rows: Alibaba Model Studio, BuildMVPFast, Atlas Cloud.

The same season in credits: three examples

Most creators buy credits, not API calls; they only become comparable as dollars per finished second.

Runway (official pricing page, 2026-09-02): Gen-4.5 costs 60 credits per 5 s. Standard is $15/month for 625 credits ($12 on annual billing), Pro $35 for 2,250, Max $95 for 9,500. That is $0.024 per credit on Standard and $0.010 on Max, so a 10-second clip is $2.88 or $1.20 — $0.29 or $0.12 per second. Standard and Pro credits reset every cycle; only Max rolls over one month.
Kling (official plan table requires login; figures from Atlas Cloud, 2026-07-14, second-hand): Standard about $6.49/month for 660 credits on annual billing; Kling 3.0 uses 45 credits per 5-second 720p clip, about $0.44 per clip or $0.09 per second. Other aggregators quote $10/month and 40–60 credits per 5 s at 1080p depending on native audio; the sources do not agree, so check the app on the day. Subscription credits expire at the end of each cycle.
SceneMixer (end-to-end: parse → reference art → storyboard → per-segment video on selectable tiers Seedance 2.0/2.5, Wan 3.0, MiniMax H3 → voice matching → compositing): published video price about $0.20–0.50 per second by tier, parsing billed per 1,000 characters, storyboards per episode, compositing per second; the mid-size pack works out near $0.011 per credit, with 500 credits free on sign-up. A 900-second season is roughly $180 at the low tier and $315–450 at 1080p/2K.

Per-second prices across the three span $0.09 to $0.50. The spread is structural: subscription platforms amortise a monthly fee, single-purpose generators exclude parsing and storyboarding, and end-to-end tools fold reference-image injection and compositing into the per-second rate. Compare total spend on the same 3-episode script, never single clips.

Per minute, per episode, per season

Band	Per minute	Per episode (90 s)	10 episodes	100 episodes
A · model only, 720p/768p, first pass	$5–17	$7–26	$70–260	$700–2,600
B · 1080p/2K list × 2.5 attempts + parsing/storyboards	$20–75	$30–115	$300–1,150	$3,000–11,500
C · small team, labor-led	$500–1,000+	$1,000 (invideo case)	$10,000	$50,000–100,000

Band A is where "a few dollars a minute" comes from. Band B is where most solo creators and small studios actually land. Band C is the invideo case study and, at the Chinese end, the ¥1,000-per-minute reports. The 100-episode column scales model cost linearly but not labor: once cast library, scene plates and templates exist, later episodes need far fewer human hours.

Five hidden costs

Regeneration. The largest item and absent from every price list. First-pass usable rates of 50–70% mean 1.5–2 calls per segment on average; crowded frames, long coats and big physical actions push past 3.
Reference-image surcharges. MiniMax H3 includes 5 reference images per call and bills $0.04 for each one beyond that (official). Three characters with a full-body sheet and a face crop each, plus one scene plate, is seven.
Tier multipliers. Runway Aleph 2.0 is 140 credits per 5 s against Gen-4.5's 60; Kling 2.5 Turbo is 15 credits per 5 s against Kling 3.0's 45 (Atlas Cloud, second-hand); Seedance 2.5 is priced 50% above 2.0 (¥42 per million tokens); Wan 3.0 "prime" is 1.5× the standard tier in Alibaba's China list. The same brand name spans a 1.5–3× range.
Expiry, minimums, rounding. Runway Standard/Pro credits reset each cycle and top-ups start at 1,000 credits; Kling subscription credits expire per cycle; ByteDance's consumer app Jimeng rounds duration up to the next step (21st Century Business Herald, 2026-04-20). Upstream download URLs expire, so budget your own bucket and egress.
Labor. invideo's case is 3 people × 3 days for 10 episodes (vendor claim). Nine person-days of review exceed the model bill at any Western day rate.

AI versus live-action, with sources

Measure	Live-action vertical drama	AI production
US series (60–80 episodes)	$15K–300K standard, $400K–600K premium; 14–21 shoot days	Band B: $300–1,150 per 10 episodes; scale ×6–8 for 60–80
US per episode	$150–3,000	$7–115 model + labor
China per minute	¥500–50,000	¥30–6,000
China per series	grassroots ¥5K–50K; mid-tier ¥500K–1.5M; S-tier ¥1.5M–5M+	entry ¥300–5,000 per 10 episodes; mid ¥10K–150K; high-end ¥200K–1M

US figures are from Streaming Lens (updated 2026-08, own survey of 65+ production companies). Chinese figures are from a Sohu comparison dated 2026-06-18. Two cautions: the "90% cheaper" claim only holds tier-for-tier — a phone-shot grassroots series and an entry-level AI series overlap; and live-action budgets are 15–25% cast and 35–45% shoot (Streaming Lens), which AI replaces with regeneration and review hours rather than eliminates.

How to cut a season's bill in half

Draft at 720p/768p, finalize at 1080p/2K. The tier gap is 1.6–2.4× and most regenerations happen at draft stage.
Push segments toward 15 s. Six segments per 90-second episode instead of twelve halves the call count; the per-second price is unchanged, the failure count is not.
Review storyboards before generating. Editing a storyboard is free; regenerating a clip is full price. Split crowded or high-motion shots first.
Stay inside free reference-image allowances. Send only the characters actually in frame.
Use promotions and free tiers for testing, not delivery. Runway's free plan is 125 one-time credits, Kling's 66 per month (second-hand), Alibaba's China region 30 seconds per Wan 3.0 tier. None covers an episode.
Budget from your own 3-episode sample, then scale linearly for model cost and sub-linearly for labor.

Key figures and sources

Figure	Value	Source	Date
Runway Gen-4.5	60 credits per 5 s; Standard $15/mo for 625 credits	runway.com	2026-09-02
MiniMax H3 API	768p $0.08/s, 2K $0.13/s; reference images beyond 5 cost $0.04 each	platform.minimax.io	2026-09-02
Kling 3.0 via resellers	$0.084–0.126/s vs $0.153/s (same model, two channels)	www.buildmvpfast.com	2026-07-01
Kling subscription (second-hand)	Standard ~$6.49/mo, 660 credits; 45 credits per 5 s 720p	www.atlascloud.ai	2026-07-14
US live-action vertical series	$15K–300K standard, $400K–600K premium; $150–3,000 per episode	lens.streaming-radar.com	2026-08-01
invideo case study (vendor claim)	$1,000 per episode, 10 episodes, 3 people, 3 days	invideo.io	2026-09-02

FAQ

Is $1,000 per episode a realistic AI short-drama cost?

It is realistic for a small team and it is the vendor's own case (invideo: 3 people, 3 days, 10 episodes). Model calls are a small fraction of that figure; the rest is labor. A solo creator using API-priced models lands at $30–115 per episode including regenerations.

How much does one AI episode cost?

For a 90-second episode: $7–45 in model calls if every clip works first time (720p to 1080p/2K), $30–115 with an average of 2.5 attempts plus parsing and storyboards, and $1,000 in a team workflow. Model cost per episode stays flat as the series grows; labor per episode falls.

Why does Kling 3.0 cost $0.084 per second in one place and $0.153 in another?

Different channels. Resellers and aggregators set their own margins, subscription credits amortise a monthly fee, and some quotes include native audio or 1080p while others do not. Name the channel and the tier before comparing; a bare "per second" figure is meaningless.

Can free credits produce a finished episode?

No. Runway's free plan is 125 one-time credits (about 10 seconds of Gen-4.5); Kling's free tier is 66 credits a month (second-hand). Enough to check that reference images and storyboards behave, not to deliver 90 seconds plus retries.

How much cheaper is AI than live action?

Per episode, US live-action verticals run $150–3,000 (Streaming Lens) against $7–115 in AI model cost plus labor. At the low end the gap is one to two orders of magnitude; a team-made AI series overlaps with grassroots live-action budgets. The saving is cast and locations, replaced by regenerations and review.

Is 720p good enough to publish?

For drafts, yes. For delivery, ReelShort- and DramaBox-style feeds recompress uploads and a 720p source visibly softens on phones. Draft low, lock the cut, then regenerate only approved segments at 1080p/2K.

Sources

Tools mentioned

Runway Kling AI Hailuo (MiniMax)Wan (Tongyi Wanxiang)SceneMixer invideo AI