The 7 Cheapest AI Video APIs Right Now, Ranked by Price Per Second

Every one of these seven ways to buy a second of AI video is ranked on one number: the price the vendor itself publishes per output second, at a stated resolution, as of September 20, 2026. This is not a "best open model" list - SLOP TV already ran that one, covering weights you can download and run yourself. This is the opposite question: what does it cost to buy a second of finished video from someone else's servers, right now, with no GPU of your own. Quality, speed and reliability are not part of the score - a slow, mediocre model that is cheap still ranks above a great one that costs more. Where a vendor bills by credit or by token instead of dollars per second, that rate has been converted using the vendor's own published example and flagged as a conversion, not a directly stated figure. Where a promotional discount is doing the work, this says so and gives the date it ends. Seven models made the cut, spanning MiniMax's H3, Black Forest Labs' FLUX.3, Alibaba's Wan 3.0, Google's Veo 3.1, Lightricks' LTX 2.5, Kuaishou's Kling 3.0 and ByteDance's Seedance 2.5 - one first-party vendor page and one hosted or optimized-infrastructure route wherever both exist.

1. MiniMax H3 (served by Pruna AI's P-Video-2-Pro)

MiniMax released H3 as an open-weight model in late July 2026 and opened it up on August 3rd, and the cheapest way to actually run it is not MiniMax's own API - it is a third-party optimization shop. Munich-and-Paris-based Pruna AI launched P-Video-2-Pro on September 17, 2026, a narrower, speed-tuned interface to H3 built around text and first/last-frame image inputs.

Pruna AI's own launch materials list $0.02 per second of finished video for 480p in speed mode, rising to $0.04 for 480p in the slower "quality" mode and $0.075 for 768p quality, according to Runtime Wire's reporting on Pruna AI's documentation and X thread. Pruna AI is also running a 50% discount on those rates through September 24, 2026, so the effective price during the promotion window is lower still. A five-second clip runs $0.10 to $0.375 depending on resolution and mode before that discount. fal.ai offers a second route to the same base model, H3 Max Turbo, at $0.025 per second at its 480p floor and $0.04 per second at 768p, according to fal's own speed comparison - close behind Pruna AI, on fal's own infrastructure rather than a third party's.

Model card

• Model: MiniMax H3 (via Pruna AI's P-Video-2-Pro)
• Maker: MiniMax, served by Pruna AI
• Price per output second: $0.02 at 480p, speed mode (50% promotional discount through September 24, 2026)
• Max clip length: 15 seconds
• Where to run it: Pruna AI API
• Try it: https://docs.api.pruna.ai/guides/models/p-video-2-pro

2. FLUX.3 (Black Forest Labs, draft mode)

Black Forest Labs made FLUX.3 Video generally available on August 4, 2026, with native audio and clips up to 20 seconds. The cheapest way into it is its own draft tier, meant for checking a composition before paying for a full render, and the two routes to it disagree on price.

On Black Forest Labs' own pricing docs, draft mode for text-to-video and image-to-video costs $0.06 per second, regardless of resolution - a third of the $0.17-per-second full HD render rate. But on fal.ai's marketplace, the same draft endpoint is billed at $0.03 per output second, half of BFL's own dashboard rate, according to Autohustle's read of fal's live endpoints on September 4, 2026. Both numbers come from the two vendors' own published rates for the same draft mode; this ranking uses the lower, fal-hosted figure, and flags the direct-API rate as the higher alternative for anyone checking a bill against BFL's own dashboard instead.

Model card

• Model: FLUX.3 (draft mode)
• Maker: Black Forest Labs
• Price per output second: $0.03 via fal.ai's marketplace endpoint; $0.06 via Black Forest Labs' own API and dashboard
• Max clip length: up to 20 seconds (full render tier)
• Where to run it: fal.ai or the Black Forest Labs API/dashboard
• Try it: https://bfl.ai/models/flux-3

3. Wan 3.0 Prime (Alibaba, via fal)

Alibaba Tongyi Lab's Wan 3.0 shipped on August 24, 2026, doubling the previous generation's clip length ceiling to 30 seconds while cutting the per-second rate in half. Fetched directly from fal's live product page, Wan 3.0 Prime bills $0.05 per output second at 480p, $0.10 at 720p, and $0.20 at 1080p for both text- and image-to-video, with the same rate holding whether the input is a prompt or a still image.

Alibaba's own direct API undercuts even that during a promotion: Alibaba Cloud Model Studio lists a 30% discount on the same tiered rate through September 24, 2026, working out to roughly $0.035 per second at 480p, according to Magica's comparison of the two routes - cheaper than fal's rate, but only until the promotion expires. Once it does, the two routes converge on the same $0.05/$0.10/$0.20 ladder. Wan 3.0's predecessor, Wan 2.7, still runs on fal for double the price - $0.10 per second - for anyone who needs its driving-audio-track input that Wan 3.0 does not carry over.

Model card

• Model: Wan 3.0 Prime
• Maker: Alibaba Tongyi Lab
• Price per output second: $0.05 at 480p, $0.10 at 720p, $0.20 at 1080p (fal); Alibaba Cloud's direct promotional rate runs about 30% below this through September 24, 2026
• Max clip length: 30 seconds
• Where to run it: fal.ai or Alibaba Cloud Model Studio
• Try it: https://fal.ai/wan-3

4. Veo 3.1 Lite/Fast/Standard (Google)

Google's own Gemini API pricing page lays out a full ladder rather than one price, and the cheapest rung ties Wan 3.0 for fourth place on this list without matching its resolution. Veo 3.1 Lite, with audio, costs $0.05 per second at 720p and $0.08 at 1080p; Veo 3.1 Fast runs $0.10 at 720p, $0.12 at 1080p and $0.30 at 4K; full Veo 3.1 Standard costs $0.40 at both 720p and 1080p and $0.60 at 4K. All of those figures are Google's own paid-tier numbers, not a reseller's.

The reason Veo lands here rather than higher: its cheapest tier is priced with audio included at 720p, one resolution step above Wan 3.0 Prime's cheapest 480p tier, and Veo's own page does not list a no-audio discount the way a reseller listing on OpenRouter does. That OpenRouter listing claims Veo 3.1 Lite without audio drops to $0.03 per second at 720p, but that figure comes from a third-party reseller rather than Google's own page, so it is noted here rather than used for the ranking.

Model card

• Model: Veo 3.1 (Lite / Fast / Standard)
• Maker: Google
• Price per output second: Lite $0.05 (720p) / $0.08 (1080p); Fast $0.10 (720p) / $0.12 (1080p) / $0.30 (4K); Standard $0.40 (720p, 1080p) / $0.60 (4K) — all with audio
• Max clip length: 8 seconds per generation
• Where to run it: Gemini API, Vertex AI, Google AI Studio
• Try it: https://ai.google.dev/gemini-api/docs/pricing

5. LTX 2.5 (Lightricks)

Lightricks' own comparison page for LTX-2.5 lists its API pricing straight: $0.09 per second at 720p, $0.13 at 1080p, $0.19 at 1440p and $0.30 at 4K - a flat resolution-based ladder with no separate tier names.

That figure sits below every route into LTX-2.5 through fal. fal's own "fastest models" write-up prices LTX-2.5 Fast at $0.13 per second at 1080p, and a separate fal comparison against Seedance 2.5 prices the slower LTX-2.5 Pro tier at $0.12 per second at 720p and $0.17 at 1080p, according to fal's own blog posts. Lightricks' direct rate is the cheaper of the two routes at every shared resolution, which is the number used for the ranking here; anyone routing through fal's Pro or Fast tiers should expect to pay somewhat more for the same open-weight model.

Model card

• Model: LTX 2.5
• Maker: Lightricks
• Price per output second: $0.09 at 720p, $0.13 at 1080p, $0.19 at 1440p, $0.30 at 4K (Lightricks' own API); fal's Pro/Fast tiers run $0.12-$0.17 at the same resolutions
• Max clip length: 10 seconds per generation on fal's Pro tier
• Where to run it: Lightricks' LTX API or fal.ai
• Try it: https://ltx.io

6. Kling 3.0 Turbo Standard (Kuaishou, via fal)

fal's live product page for Kling Video V3 Standard Turbo bills $0.112 per output second for 720p video with native audio - a five-second clip runs $0.56. That is the cheapest of Kling's tiers on fal; the plain Kling 3.0 Standard tier costs $0.126 per second with native audio for the same 720p output ($0.084 without audio), no cheaper despite the lower tier name, and Kling's Native 4K tier runs $0.42 per second, roughly triple the category rate.

The $0.112 figure holds up against Kling's own pricing structure, too. Kuaishou's Kling bills in credits rather than dollars directly - 6 credits per second at 720p without audio, rising to 12 credits per second at 1080p with native audio, according to Kling's official credit guide. Converted to dollars using Kling's own credit value, Magic Hour's pricing breakdown, checked September 13, 2026, puts Kling 3.0 Turbo with native audio at $0.112 per second at 720p and $0.140 at 1080p - matching fal's hosted rate almost exactly, which suggests fal is passing Kling's own pricing through rather than marking it up.

Model card

• Model: Kling 3.0 Turbo Standard
• Maker: Kuaishou
• Price per output second: $0.112 at 720p with native audio (fal.ai, matching Kling's own credit pricing as converted by Magic Hour)
• Max clip length: 10 seconds per generation
• Where to run it: fal.ai or Kling's own API
• Try it: https://fal.ai/models/fal-ai/kling-video/v3/turbo/standard/image-to-video

7. Seedance 2.5 (ByteDance, via fal)

ByteDance's Seedance 2.5 is the most expensive model that still clears the bar for this list: a real, vendor-published price per output second. fal's own product page for Seedance 2.5 image-to-video states the model bills by token - $0.0214 per 1,000 tokens at 480p and 720p - and works that out to roughly $0.2205 per second at 480p and $0.4730 per second at 720p for standard 16:9 output, both figures fal publishes directly on the same page.

ByteDance's own BytePlus billing, upstream of fal, uses the same token structure but a different headline rate: $10.70 per million tokens without a video-reference input. Run through the same formula, that works out to about $0.103 per second at 480p and $0.231 per second at 720p, according to Seedance2.so's math applied to BytePlus's official rate card - roughly half of what fal charges for the same output. Because that lower number is a derived conversion rather than a rate BytePlus states directly in dollars per second, this ranking uses fal's directly published $0.2205-per-second figure, and flags the cheaper BytePlus-derived route as the one to check before committing to a vendor. Either way, Seedance 2.5's real advantage is not price - it is the single 30-second take and up to 50 reference inputs neither of the other six models here can match.

Model card

• Model: Seedance 2.5
• Maker: ByteDance
• Price per output second: $0.2205 at 480p, $0.4730 at 720p (fal.ai's directly published rate); BytePlus's own token rate converts to roughly $0.103-$0.231 at the same resolutions
• Max clip length: 30 seconds
• Where to run it: fal.ai or BytePlus/Volcengine's own API
• Try it: https://fal.ai/models/bytedance/seedance-2.5/image-to-video