VoidSource Media Cost IndexRoutes and subscriptions checked Jun 22, 2026–Sep 4, 2026 · Quality ratings Sep 4, 2026
What AI video actually costs.
Video tools price by second, clip, credit wallet, subscription tier, and sometimes by route. This hub normalizes the moving parts into comparable prices, then pairs them with quality signals so the cheap option and the useful option are not confused. For agents and automated filmmaking, price is only half the question; the sweet spot is a route that is both cheap and callable without a human in the loop, so metered API routes and human-first credit plans stay separate here.
The short answer
Across all 643 tracked USD metered video API routes in this comparison, a 10-second clip currently ranges from $0.05 (P-Video via Replicate) to $20.80 (Seedance 2.5 via OpenRouter). The table below previews the cheapest rows; credit subscriptions can beat API pricing at high utilization, but at the default 40% usage assumption, unused credits make the sticker rate materially too low.
16
pricing sources
880
comparable route prices
$0.005-$2.08
USD per-second API range
165
credit plans decoded
The verdict
Gemini Omni Flash is the highest-rated video model right now.
Best value (top quality at each price point): P-Video, Seedance V1.5 Pro, Grok Imagine Video 720p, Veo 3.1 Fast Audio. Best open-weights: Kandinsky 5.0 T2V Pro.
01
Where the sweet spots are
The short answers, before the evidence. Each verdict is derived from the prices and quality ratings below, and each links to the section that backs it. Where no public quality rating exists, we say so instead of pretending.
PrototypingCheapest current-gen drafts?
Seedance 2.0 Minino public Elo
$0.036/s480p first-party (ModelArk) · $0.076/s at 720p
The cheapest current-generation ByteDance route we track. No arena has rated it yet, so there is no quality signal to set against the price — cheap here means cheap, not proven.
$0.152/s720p first-party (ModelArk) · $0.22/s via Replicate
Tops both arena boards among models with a tracked metered price. Gemini Omni Flash rates higher on text-to-video, but no API price we can verify exists for it.
What the pricing data already shows: most routes on the board are metered APIs an agent can drive, and ModelArk doesn't bill failed generations. What it does not yet show is setup friction, async job handling, retry behaviour, or cost per clip you actually keep — so there is no ranking here yet.
Blind human-preference Elo is useful by itself, but the real decision starts when it is joined to price. Up and left is better; the frontier marks models where nothing stronger costs less.
Arena Elo Sep 4, 2026 · USD per second where tracked
Open weightsClosed / APIValue frontier
Every rated model, joined to every tracked route and sortable by Elo or price, lives in the board below.
Rated on this board, price untracked
Gemini Omni Flash Elo 1512Wan2.7 T2V Elo 1344Wan2.6 T2V Elo 1329Hailuo 2.3 Elo 1205Ray 3 Elo 1205Kandinsky 5.0 T2V Pro Elo 1172Hunyuan Video 1.5 Elo 1169LTX 2 19b Elo 1152Kandinsky 5.0 T2V Lite Elo 1113Sora Elo 1069Ray2 Elo 1064Pika V2.2 Elo 1008
These models hold an Arena rating but expose no clean public API price we track yet. We show the gap instead of estimating it.
Quality leader
Gemini Omni Flash leads this board at Elo 1512, but public API pricing is not tracked yet.
Priced quality
Dreamina Seedance 2.0 720p is the highest-rated model with a tracked price in this board.
Efficiency corner
P-Video is the cheapest priced board point at $0.02/sec.
03
Same model. Same settings. Different price.
Wrappers and marketplaces often resell the same underlying video model. When the comparability key matches, we can show the spread without hand-waving: same model, operation, resolution, tier, audio mode, currency, and execution mode. Every seller in the group is marked on the bar, not just the endpoints.
Seedance 2
video · 4K · default
$0.467 · OpenRouter$1.25 · Replicate
2.7xspread
Seedance 2
video · 1080p · default · 3 providers
$0.228 · OpenRouter$0.569 · ModelArk
between: $0.45 · Replicate
2.5xspread
Happy Horse
video · 1080p · default
$0.128 · OpenRouter$0.28 · Replicate
2.2xspread
Vidu Q3
video · 540p · pro
$0.045 · Vidu$0.07 · Replicate
1.56xspread
Vidu Q3
video · 720p · pro
$0.1 · Vidu$0.15 · Replicate
1.50xspread
Happy Horse
video · 720p · default
$0.099 · OpenRouter$0.14 · Replicate
1.42xspread
The useful signal is not only the largest markup. A near-1.0x row is just as valuable: it means the marketplace is effectively pegged, so convenience may be worth taking. A high-spread row is where teams should check the direct API first. Across all 18 comparable multi-provider groups we track right now, Replicate is the most expensive seller in 13 of them.
04
The board: every model, every route
107 video models joined to all 643 tracked USD-metered API routes. Sort by Arena Elo or price; expand a model to see every provider selling it: same model, same resolution, different bill. Credit-plan economics are handled separately below.
Model rows show each family's cheapest fully-specified metered route; click a row to see every tracked route with operation, tier, audio, and the provider's own pricing page as evidence. Models with an Arena rating but no clean tracked route show their board-level price instead, labeled by source; prices from flagged extractions are tagged unverified and never pose as clean data. Subscription and credit routes are intentionally separated below because utilization and credit expiry change the effective cost.
The market at a glance · $/sec by resolution
log scale · ▲ median
360P7 routes
▲
$0.025 · $0.07 · $0.1
480P79 routes
▲
$0.008 · $0.05 · $0.103
540P17 routes
▲
$0.02 · $0.07 · $0.144
720P171 routes
▲
$0.005 · $0.1 · $0.41
1080P159 routes
▲
$0.01 · $0.15 · $0.72
4K37 routes
▲
$0.163 · $0.4 · $2.08
n/a94 routes
▲
$0.017 · $0.128 · $0.5
Every tracked route counts here, including flagged extractions: min · median · max per resolution. Same model families appear at several points; expand a model above to see why.
Official model weights
Repository links for the exact model version. Check the source for access terms and setup requirements.
Alibaba’s video model, offered through hosted generation endpoints.
Sources and limitations
Date is documented availability on fal, not proof of Alibaba’s first model release. The reported earlier beta and financial news are not part of this event.
Pick a volume, a clip length, and (honestly) how many generations you pay for per clip you keep. Totals are computed over the same clean tracked routes as the board; flagged extractions and vendor draft modes never enter a total.
Floor estimates: most providers bill fixed duration blocks (a 7-second request often bills as 10 seconds) and cap clips around 10–15 seconds, which is why lengths above 12s are not offered. Reference-video operations (video as input) appear as their own, pricier routes. Prices exclude taxes and volume discounts.
06
Prototyping economics: pay draft prices for draft work
Iteration is where video budgets actually go: you generate ten clips to keep one. ByteDance's ModelArk terms include three levers that cut what a draft costs, none of which show up on a per-second comparison chart.
Draft mode
Seedance 1.5 Pro generates low-quality test drafts at a reduced token rate: 480p only, and currently the only model that supports it.
Offline inference
50% offthe token rate
Batch jobs that can wait pay half: Seedance 1.5 Pro, 1.0 Pro, and 1.0 Pro Fast all list an offline rate at exactly 50% of instant. The catch: the whole Seedance 2.0 line is “not supported yet”, so if you prototype on 2.0 you pay full rate.
Failure billing
$0for failed generations
ModelArk only charges for successfully generated videos; failures and content-moderation blocks cost nothing. For an agent pipeline that retries, this changes the effective cost per accepted clip.
These are the vendor’s own estimates at standard specs (16:9); actual billing follows the token meter. We have not verified that Google’s batch-inference discount covers Veo, so we don’t claim it here. BytePlus ModelArk pricing
07
The subscription illusion
Credit plans are not wrong, but their sticker price assumes every credit gets used. Pick how much of a monthly allowance you would realistically burn; unused credits raise the effective price per credit before any model-specific burn rate is applied.
Credits you actually use:
advertisedeffective at 40% use
Higgsfield · Ultra
$209/mo · 7,200 cr
$0.029
$0.073 · 2.5x
Hailuo · Max
$200/mo · 20,000 cr
$0.01
$0.025 · 2.5x
Dreamina · Advanced
€52/mo · 7,955 cr
€0.0065
€0.016 · 2.5x
Adobe Firefly · Premium
€220/mo · 50,000 cr
€0.0044
€0.011 · 2.5x
Freepik · Pro
$280/mo · 333,333 cr
$0.0008
$0.0021 · 2.5x
The API question and the subscription question are different. APIs win when you need automation, repeatability, auditability, or agent integration. Credit plans can win when a wrapper provides workflow value, API access is missing, or you reliably burn the monthly allowance.
The flip point is utilization: a plan that looks cheaper at 100% usage can become materially more expensive once credits expire.
Wrapper vs API, exact matches at 40% use
Seedance 2 Fast
Higgsfield Ultra vs OpenRouter
4.6x
Never beats the OpenRouter route: pricier even at 100% use.
Seedance 2
Higgsfield Ultra vs OpenRouter
3.4x
Never beats the OpenRouter route: pricier even at 100% use.
Seedance 2
Higgsfield Ultra vs OpenRouter
3.4x
Never beats the OpenRouter route: pricier even at 100% use.
Automation access matters
For an automated pipeline, the sweet spot is the cheapest route an agent can actually drive: a metered API with async jobs, retries, and usage accounting, which is most of the route board above. Credit subscriptions are mostly human-first by design: plan credits buy clicks in a web UI, not API calls. A wrapper can still be the right choice when its product layer is the workflow, and the line is blurring: Higgsfield exposes agent access via MCP and CLI billed in plan credits, and Kling now does too (first-party MCP server and CLI at the same credit prices as its apps, but agents get no off-peak free generations and cannot spend bonus credits, so an agent's credit buys less than a human's). The economics above apply either way: when the meter is credits, utilization and expiry set your real price; when it is a per-second API, the listed rate is what you pay.
Prices are captured from provider pricing pages or APIs, then normalized to per-second or per-credit terms. The route table links each listed API price to its source. Comparisons keep operation, resolution, tier, audio, currency, and region separate, so a near miss is never passed off as the same route.
Route and subscription prices were last confirmed between Jun 22, 2026 and Sep 4, 2026. The route table links each listed API price to its source.
How much does AI video cost?+-
This page reports USD metered API costs per second, converts them into 10-second clip costs, and separates subscription pricing so unused credits do not get hidden inside one number. The route table links each listed API price to its source.
Why can the same model have different prices?+-
Wrappers and marketplaces can resell the same model with their own margins, convenience layers, or plan economics. We only publish a spread when the model, operation, resolution, tier, audio mode, currency, and execution mode match.
Is a subscription cheaper than an API?+-
Sometimes. A subscription can win when you use most of the monthly credits or need the wrapper workflow. APIs usually win when you need automation, agent access, repeatable jobs, and no use-it-or-lose-it credit decay.
Can AI agents generate videos?+-
Yes, through two doors. Metered APIs are the agent-native door: a key, an async job queue, and a per-second or per-generation price (BytePlus, Google Veo, Kling's developer platform, MiniMax, fal, Replicate). MCP is the second door: an agent drives the tool through a standard protocol instead of custom integration code. What an agent cannot use is a human-first credit subscription: plan credits that only buy clicks in a web UI.
Which AI video tools support MCP?+-
First-party MCP support is still rare but growing. MiniMax publishes an official MCP server for its video, image, and speech models, billed like normal API usage. Higgsfield exposes MCP and CLI access billed in the same credits as its subscription plans. Kling now ships both a first-party MCP server (kling.ai/mcp) and a CLI, billed in the same credits as its consumer plans, with two agent-specific catches: off-peak free generation is not available through assistants, and bonus credits cannot be spent there, so an agent's credit buys strictly less than a human's. Most other providers are reachable by agents only through their metered APIs, or through third-party MCP wrappers that put someone else's code and margin between you and the meter.
Are AI video credit plans usable for automation?+-
Mostly no, by design. Consumer credit plans bill a human clicking a web UI; where the same company sells developer access, it is a separate metered product with separate pricing (Kling membership vs. Kling's Open Platform, Dreamina vs. ByteDance's ModelArk). The exceptions among the credit platforms tracked here are Higgsfield and, since mid-2026, Kling; both expose agent access via MCP and CLI on plan credits. Neither hands agents the full plan value: Higgsfield's Unlimited mode stays website-only, and Kling excludes off-peak free generation and bonus credits from assistant use. Agent-billed credits inherit the same utilization and expiry economics measured on this page.
Is MCP better than an API for AI video generation?+-
They solve different problems. A metered API is the cleaner primitive for production pipelines: queues, retries, webhooks, exact usage accounting. MCP removes integration code so an agent can drive the tool directly. The economics matter more than the protocol: MCP calls billed against an API cost the listed rate, while MCP calls billed in subscription credits inherit utilization and expiry, the credit breakage quantified above.
Why are some top video models missing from the price table?+-
Quality boards and pricing sources do not update on the same cadence, and some high-quality systems do not expose clean public API pricing. Those gaps should stay visible rather than being filled with estimates.