A five-second AI video clip is cheap per attempt. Across the documented API lineup — roughly $0.02 to $0.60 per output second — one attempt at five seconds costs about $0.10 to $3.00, and about $0.25 to $2.00 on the mainstream 720p tiers; simple multiplication of cited rates, not a quote [2][3]. But the sticker is nearly useless for budgeting. What you actually pay for is a usable second, and retries, minimum charges and format surcharges multiply the bill. China's comic-drama studios mass-produce anyway, and not because inference became free — the savings are organizational, and the money that decides the margin goes to traffic buying, the one cost generation has nothing to do with.
So the number worth carrying is not what one clip costs but what one deliverable costs. Fix your resolution and audio requirements first, then budget in billed seconds multiplied by a realistic retry count.
The sticker price is low and specific
Runway's official API pricing page is the cleanest baseline: one credit is $0.01, and every video model bills per second of output at the resolution you choose [1]. On that list, Gen-4 Turbo runs $0.05 per second, Gen-4.5 $0.12, and Veo 3.1 with audio $0.40 — against $0.02 for Grok Imagine 1.5 Lite at 480p [2]. An eight-second clip therefore spans $0.16 to $3.20 across the documented lineup [2]. At those per-second rates, a five-second attempt runs $0.10 on the cheapest tier, $0.25 on Gen-4 Turbo, and $2.00 on Veo 3.1 with audio — the same multiplication, per tier.
Google's Veo 3.1 tiers — retired with the preview family in October — were priced per output "count" with no published count-to-seconds conversion; the per-second readings of those tiers (Lite from $0.05 at 720p to Standard's $0.60 at 4K) are the vendor comparators' conversion, resting on the alignment between Google's count prices and Runway's per-second resale rates — an inference czlytech labels as such [2][3].
The rates also move: Runway's page currently lists Kling 4.0 at a promo $0.1256 per second in 720p (list $0.157), 20% off through November 30, 2026 [1]. And the purchase channel changes the number as much as the model does: MiniMax sells a $1,000 Hailuo video pack of 3,760 points valid for one month — roughly $0.029–$0.089 per second, but only if the pack is fully consumed before it expires [3]. Resolution, audio and purchase channel are the three conditions to check before repeating any of these figures.
The bill is not the sticker
The pricing explainers that track these pages converge on one rule: the cost of a usable clip equals the unit price times your retry count, and the four things that blow past a per-second estimate are retry rate, billing block size, input video billed separately, and 4K/HDR multipliers [2]. Runway's page shows the mechanics. For Kling 4.0 jobs that include input video, Runway lists a separate rate — 9.4 credits ($0.094) per second at 720p before the promo — and the comparators flag that input and reference video may be billed separately on several models [1][2]. Output-format surcharges stack on the model's rate for Gen-4.5 and Aleph 2.0 output: +$0.05 per second for ProRes or PNG sequences, +$0.20 for 10-bit SDR below 4K, +$0.20 for HDR on Gen-4.5, and +$0.40 for 10-bit SDR at 4K [1]. Minimum charges bite when clips are short: Seedance 2.0 carries a 56-credit minimum per request [2], and on Runway even audio generations have floors and round up to whole credits [1].
One illustration — the source's own editor calculation, not a quote — puts it together: a 30-second film from four 8-second shots retried three times is roughly 96 billed seconds, or $4.80 to $38.40 of pure generation cost, before editing, voice and licensing [2]. There is little relief upstream either: mainstream video APIs essentially have no free tier, and consumer free credits do not transfer to API balances [4]. The one mercy is failure handling — MiniMax does not charge for review-failed tasks, and Seedance's official channel promises no charge for failed generations [4], which matters when retries are the budget killer.
So if generation still costs this much per usable second, how do comic dramas mass-produce?
The savings are organizational, not free physics
The China-market numbers, in RMB as reported: a premium live-action paid short drama averages ¥1.5 million to produce, while a vertical AI pseudo-real drama can be held under ¥200,000 — some titles reportedly as low as ¥20,000–30,000 — and a 90-minute comic drama costs about 80% less than the traditional route [5]. The production line's real description, per the same report, is "super-individual" output: one person, one day, one drama [5]. Kunlun Wanwei says its per-drama cost fell from about US$200,000 to about US$20,000; CN Online says AI short-drama cost per minute runs more than 75% below traditional 3D animation — both company-reported figures [5].
On this record, the savings are organizational rather than physical: template-driven production lines need fewer people-hours per finished drama, and the models on those lines list at the same kind of public per-second rates as any API customer's. Seedance 2.5, ByteDance's video model, lists at about $0.20–$0.23 per second at 720p across reseller tables — one reseller's published rate, with the official token-based rate cited alongside [4] — and Runway's own documentation lists Seedance 2.5 among its billable video SKUs at four resolutions [6]. A single Seedance 2.5 request can generate up to 30 seconds [4]. Runway itself is also selling deliverable pricing: its Product Ad recipe offers a finished 4-second 720p ad for a flat $1.76 plus $0.30 per additional second [1]. And voice is a rounding error next to video seconds — Seed Audio 1.0 bills $0.0025 per second, Eleven v4 $0.05 per 1,000 characters of script [1].
Traffic, not generation, decides whether these dramas make money
The money goes to traffic. In China's short-drama chain, traffic buying is roughly 70% of total cost; IP licensing runs from tens of thousands to hundreds of thousands of RMB per title; and cloud AI compute prices rose by as much as 34% in April [5]. Lin Qiwen, vice president of DataEye's Juanchacha drama-data platform, puts the squeeze in one sentence: production cost down about 90%, traffic cost up more than 100% year over year, revenue down more than 50% [5]. Per 1,000 plays, revenue has fallen from about ¥60 in late 2025 to ¥15–30 in 2026, per Securities Daily via the same report, and mature media-buying teams hold ROI at a thin 1.03–1.07 [5].
The volume numbers explain why. Douyin added about 221,900 AI dramas in the first half of the year; in the same DataEye-based report, only 0.47% of AI short dramas count as hits, and comic dramas are harsher — one winner per 1,000 new titles, under 0.1% [5]. The China Netcasting Services Association's data shows AI dramas were 95% of first-quarter launches but drew only 4% of live-action playback volume [5]. For scale: DataEye puts the market above ¥22 billion for January–May, eyeing ¥40 billion for the year, with users past 600 million [5]. Platforms are pushing back on the oversupply — Hongshu delists low-quality AI dramas and cancelled minimum guarantees for small and mid-team AI pseudo-real dramas, keeping roughly a 20% revenue split, and Douyin cut the pseudo-real drama split coefficient from 60% to 40% [5]. The tool vendors are the clearest winners so far: Kuaishou's Kling reports Q1 revenue above ¥650 million, up over 300% [5].
Budget in billed seconds; check the market state
For the budget, the rule is the synthesis of both halves: estimate a deliverable as billed seconds × realistic retries × your tier's per-second rate, check whether input video bills separately on your provider and count it when it does, add format surcharges when they apply, and count minimum charges on any short job [1][2]. On the documented lineup, a five-second attempt spans about $0.10–$3.00 — about $0.25–$2.00 on the mainstream 720p tiers; anything longer is a budget you build, not a price you look up.
The market state matters as much as the arithmetic, because two former price anchors are gone. OpenAI removed Sora 2's API on September 24, 2026, with no official replacement, per two API-market trackers [4][3], and Google closed the Veo 3.1 preview family on October 22, pointing to Gemini Omni Flash — about $0.10 per second at 720p on token-based billing [4][3]. Any budget built on old Sora or Veo numbers needs rebuilding. Microsoft's Azure route prices Sora 2 per second — $0.10/$0.30/$0.50 for Global-standard/Pro/Pro High Res, with Data Zone deployments about 10% higher, per czlytech's reading of Azure's rates [2]; Azure's own pricing page confirms the Global versus Data Zone deployment split [7]. It is a preview with limits that shape rework: 4/8/12-second API parameters, at most two 720p variants per prompt, two concurrent tasks, and tasks that expire after 24 hours [2].
On the supply side that mass production feeds, releases now run through policy gates: AI comic dramas have required filing before release since April, classification standards landed in June, and a four-month minor-protection campaign started in July [5] — while funding keeps flowing into the production lines, from JingYing Technology's tens of millions of US dollars to AniShort's nearly ¥100 million [5]. Per attempt, AI video is cheap and public. Per deliverable, it is a budget — and the studios that mass-produce are the ones who counted retries and traffic before they counted seconds.