AI Video Model Pricing (Aug 2026): Official Per-Second Rates, Normalized
Last updated August 7, 2026
AI video pricing spans five billing units — per second, credits, tokens, per video, and metered inputs — and official per-second rates run $0.02 (Pruna) to $0.70 (Sora 2 Pro), a 35x spread. Normalized to a 10-second 1080p clip with audio, costs range $0.40 to $8.40; MiniMax-H3 delivers 2K at $1.30. Reseller quotes run 2-3x over official rates, so this page cites vendor rate cards only.
Updated August 2026
Ask what an AI video model charges per second, and fewer than half the market can answer in those units. Veo 3.1 bills per second of output. Kling sells subscription credits, with exactly one yuan-denominated per-second rate on the API side. Seedance bills per million tokens, the way language models do. Hailuo prices its older models per finished video and its newest per second — including per-second billing on your input references. PixVerse charges credits per second that shift with resolution and an audio toggle. So this page does the only honest comparison available: it pushes every official rate through the same job — a 10-second, 1080p clip with sound. One ground rule: every figure is the vendor's own, as of August 2026, because third-party resellers routinely quote per-second prices marked up 2–3× over official rates — and their numbers contradict each other. Where no official price exists, this page says so rather than borrowing one.
Which models publish true per-second rates?
Seven families do, spanning official dollar and yuan rate cards:
| Model | Official rate (Aug 2026) | Audio | Source |
|---|---|---|---|
| Veo 3.1 | $0.40/s (720p–1080p), $0.60/s (4K) | Always on, included | Google Gemini API pricing |
| Veo 3.1 Fast | $0.10/s (720p), $0.12/s (1080p), $0.30/s (4K) | Always on, included | Google Gemini API pricing |
| Veo 3.1 Lite | $0.05/s (720p), $0.08/s (1080p) | Always on, included | Google Gemini API pricing |
| Sora 2 API | $0.10/s (sora-2, 720p) to $0.70/s (sora-2-pro, 1080p) |
Native | OpenAI model page — API sunsets Sept 24, 2026 |
| Kling 3.0 Turbo | ¥0.8/s (~$0.11) at 720p — the family's only official per-second rate | Native since 2.6 | Kling API |
| MiniMax-H3 (Hailuo) | $0.13/s — at 2K | Native stereo | MiniMax Open Platform |
| HappyHorse | ¥0.9/s (~$0.125) at 720p; ¥1.6/s (~$0.22) at 1080p | Joint single-pass A/V | Alibaba Cloud Model Studio (Bailian) |
| LTX-2.3 | Fast $0.06–0.24/s, Pro $0.08–0.32/s (1080p → 4K) | Included | ltx.io API pricing |
| Grok Imagine Video | 1.0 $0.05/s; 1.5 $0.08/s (up to 1080p) | One-pass | xAI docs |
| P-Video (Pruna) | $0.02/s (720p), $0.04/s (1080p) | Audio output included | Pruna launch announcement |
| FLUX 3 | No official pricing — Early Access only | Native | Black Forest Labs |
Official per-second pricing spans $0.02 to $0.70 — a 35× spread — and the extremes are at least honest about why: Pruna sells iteration speed below frontier fidelity, while Sora 2 Pro was priced as a flagship and is now a sunsetting one. The row worth staring at is MiniMax-H3: $0.13 per second buys 2K output — higher resolution than anything else on the table — at roughly a third of Veo 3.1's 1080p rate. And the FLUX 3 row is a warning in miniature: any FLUX 3 price you see quoted today is a reseller's invention, because Black Forest Labs has published none.
How does Seedance's token billing work?
Seedance is the outlier that explains why reseller numbers conflict so badly. ByteDance bills it the way LLMs are billed: the video you request is metered in tokens, and you pay per million of them. On Volcano Engine the official rates are ¥28 per million tokens for requests with video input and ¥46 per million for pure text-to-video; ByteDance's international platform (BytePlus) lists 720p generation at $7.00 per million tokens without video input and $4.30 per million with it. Token count scales with resolution and duration together — a longer, sharper clip is simply more tokens — and the official rate math works out to roughly ¥1 (~$0.14) per second of output at typical 1080p-class settings, with Seedance Mini starting around ¥0.16/s (~$0.022). Treat that per-second figure as a planning approximation, not a price: the invoice is denominated in tokens.
This is also where the 2–3× markup problem is worst: resellers convert token rates into per-second price lists using undisclosed assumptions, which is why no two of their Seedance price pages agree. If a quoted rate can't be traced back to ¥28–46 per million tokens, it isn't official.
What makes Hailuo's billing asymmetric?
MiniMax prices the Hailuo line along two axes at once. The pre-H3 models bill per finished video: officially $0.28 for 6 seconds at 768p, $0.49 for 6 seconds at 1080p, $0.56 for 10 seconds at 768p — plus a little-known 512p tier around $0.08 per clip. MiniMax-H3 (July 31, 2026) switches to $0.13 per second at 2K, then meters the inputs: reference videos billed per second of input footage, reference images free for the first 5 and $0.04 each after, audio references free. Failed generations aren't billed.
The verdict on that structure: MiniMax charges for what is expensive to process and gives away what isn't. If your workflow leans on audio references, H3 is effectively discounting it; if you feed it long reference videos, your "per-second" cost quietly includes seconds you supplied, not just seconds you got back.
What do Kling and PixVerse credits actually buy?
Kling is subscription-first: a free tier of 66 credits per day, then Standard at $10, Pro at $37, Premier at $92, and Ultra at $180 per month, with Kling 2.5 Turbo priced at 25 credits per 5-second clip. The family's only official per-second rate is ¥0.8/s (~$0.11) for Kling 3.0 Turbo at 720p; 1080p runs through plan credits, so its effective cost depends on your tier. Partner platforms publish their own per-second Kling rates — partner prices, not Kling's. Video extension, note, is a paid-plan feature.
PixVerse publishes an unusually clean credit-per-second table in its platform docs. V6 runs from 5 credits/s at 360p without audio to 18 credits/s at 1080p, or 23 with audio on — the audio toggle adds about 28% at 1080p. C1, the film-production model, runs one credit higher across the board (19/24 at 1080p). A 10-second 1080p clip with audio is therefore 230 credits on V6. What PixVerse does not publish is an official dollar value per credit, which is why it appears in credits — not dollars — in the table below.
What does a 10-second 1080p clip with audio cost on each model?
Three caveats, because three vendors can't sell exactly this clip: Veo 3.1 caps a generation at 8 seconds (and 1080p is 8s-only), Sora's API durations step 4/8/12/16/20 seconds so its row prices the 12-second minimum that covers 10, and Kling's only official per-second rate is 720p.
| Model | The 10s/1080p/audio bill | The catch |
|---|---|---|
| P-Video (Pruna) | $0.40 | Draft-tier fidelity by design |
| LTX-2.3 Fast | $0.60 | — |
| LTX-2.3 Pro | $0.80 | Open weights, if you'd rather self-host |
| Grok Video 1.5 | $0.80 | Dialogue uses preset voices only |
| Veo 3.1 Fast | $0.96 — for 8s at $0.12/s | 8-second generation cap |
| Kling 3.0 Turbo | ~¥8 (~$1.10) — at 720p | 1080p runs on plan credits instead |
| MiniMax-H3 | $1.30 — and it's 2K, not 1080p | Input references billed separately |
| Seedance 2.x | ≈¥10 (~$1.40), token-billed approximation | Invoiced in tokens, not seconds |
| HappyHorse | ¥16 (~$2.20) | 15-second generation ceiling |
| PixVerse V6 | 230 credits | No official credit-to-dollar rate |
| Veo 3.1 | $3.20 — for 8s at $0.40/s | 1080p is 8s-only |
| Sora 2 Pro | $8.40 — for 12s at $0.70/s | API shuts down Sept 24, 2026 |
Same ten seconds, roughly a 17× spread in per-second rates from bottom to top — but the ranking isn't the insight. The insight is where resolution-per-dollar peaks: in the middle, at MiniMax-H3, which returns 2K frames for less than half of what Veo 3.1 charges for two fewer seconds of 1080p. Above the ~$2 line you are paying for something other than pixels — Veo's always-on audio, 4K tier, and Google-ecosystem integration, or in Sora's case, a model with seven weeks to live.
The number the rate card hides: retries
Sticker price per second understates real cost because prompting video is trial-and-error — PixVerse's own review of V6 calls it "not a magic one-try-solves-everything model," and that candor applies across the market. Your effective rate is the official rate multiplied by attempts per keeper, which is why the cheap-draft mechanisms matter as much as the rates: Pruna's 2.5-second draft mode, Seedance 1.5's Draft Sample preview (a 65% efficiency gain, per ByteDance), a 21-credit 3-second PixVerse test before a 345-credit commit — and MiniMax and HappyHorse don't bill failed generations at all. Price the iteration loop, not the clip.
AI video pricing questions
What's the cheapest AI video model with official pricing? Pruna's P-Video: $0.02/s (720p) and $0.04/s (1080p), audio included, positioned deliberately below frontier quality. Among frontier-adjacent models the floor is LTX-2.3 Fast at $0.06/s (1080p) — with open weights if you'd rather remove per-second cost entirely.
Why do reseller prices differ so much from these numbers? Wrapper sites resell API access at a margin — typically 2–3× over official rates — and their price pages contradict each other, worst of all for token-billed Seedance. Price against the vendor rate cards cited here or not at all.
How much does Veo 3.1 cost per second? Per Google's Gemini API pricing as of August 2026: $0.40/s at 720p–1080p and $0.60/s at 4K, with Fast at $0.10/s (720p) and Lite at $0.05/s (720p). Audio is always on and included in the rate.
Is Sora 2 still worth including in a price comparison? Only until September 24, 2026, when OpenAI's API sunset takes effect. After that date there is no legitimate way to buy Sora generations at any price.
What does FLUX 3 cost? Nothing official has been published as of August 2026 — the model is in Early Access. Until Black Forest Labs posts rates, every quoted FLUX 3 price is fabricated.
Comparing rates is easier when the models share a roster
Acting on an 11-way price comparison normally means eleven accounts, four currencies, and three billing units. Every model on this page runs inside invideo, the AI video platform that gives serious creatives every major model in one place — which turns pricing into a per-shot choice rather than a procurement decision: draft on the cheap tier, spend frontier rates on the shots that earn them. The lineup is browsable at invideo.io/ai-models; the AI video generator workflow is where pick-per-shot plays out.
Version history: first published August 2026 with official rates current as of that month — Veo 3.1/Fast/Lite, Sora 2 (pre-sunset), Kling 3.0 Turbo, Seedance token rates, MiniMax-H3, HappyHorse, PixVerse V6/C1 credit tables, LTX-2.3, Grok Video 1.5, and Pruna P-Video. Rates change; this page tracks the official cards, not reseller quotes.