Subscribe
video generation
5 min read

What a Second of AI Video Actually Costs in Late 2026

Normalized per-second pricing across Sora's post-shutdown market: from $0.015 on aggregators to $1.50 at 4K, with the quality leaderboard matching the price board even less than before.

Retro editorial illustration: a balance scale weighing a strip of film against a stack of coins, with a filmstrip of price tags and coins curving across the frame
Retro editorial illustration: a balance scale weighing a strip of film against a stack of coins, with a filmstrip of price tags and coins curving across the frame

What a Second of AI Video Actually Costs in Late 2026

The AI video API market spent 2026 consolidating around a single billing unit, the output second, and then quietly made the price impossible to compare. The spread between the cheapest and the most expensive way to render one second of video is now roughly 100 to 1, and the ranking of models by price has almost nothing to do with the ranking by quality. Here is what the published rate cards actually say, normalized to dollars per second.

First, the gap in the table: Sora is gone

OpenAI shut the consumer Sora web and app experiences on April 26, 2026, and removed the Videos API and every Sora 2 model alias (sora-2, sora-2-pro and their dated snapshots) from the API on September 24, 2026, with no successor model listed in the deprecation documentation. At list prices before removal, sora-2 ran $0.10 per second at 720p and sora-2-pro ran $0.70 per second at 1080p. OpenAI has said Sora continues as a research project on world models while compute is redirected toward coding tools and enterprise products. The practical effect for buyers: one of the two reference points every 2025 comparison chart was built around no longer sells anything.

Bars of very different heights, one tiny and one towering, on a platform; a hand drops a coin into the tall one

The rate cards, normalized

Three billing conventions are in use. Runway's developer API charges in credits at a fixed $0.01 per credit. Google's Gemini API charges video output in tokens (5,792 tokens per second of 720p output). Inference hosts like fal charge per second directly. All figures below are converted to dollars per output second at the stated resolution.

Budget tier (silent or low-res):

  • MiniMax H3 Max Turbo, text-to-video via fal.ai: $0.015/s
  • Grok Imagine 1.5 Lite at 480p via Runway Dev: $0.02/s
  • MiniMax H3 Max, text-to-video via fal.ai: $0.03/s
  • Runway Gen-4 Turbo: $0.05/s
  • Google Gemini Omni 1.2 Flash, direct on the Gemini API: about $0.07/s at 720p ($12 per million output video tokens; $0.023/s at 360p)

Mid tier (1080p class with audio):

  • Alibaba Wan 3.0 at 720p via Runway Dev: $0.10/s; at 1080p: $0.20/s
  • Gemini Omni 1.1 Flash at 720p via Runway Dev: $0.10/s
  • Runway Gen-4.5: $0.12/s
  • Kling 3.0 Pro at 1080p via fal.ai: $0.14/s; via Runway Dev with audio: $0.17/s

Premium tier (4K and full-flagship):

  • Veo 3.1 with audio via Runway Dev: $0.40/s
  • Kling 3.0 4K via Runway Dev: $0.43/s
  • Gemini Omni 1.2 Flash with 4K upsampling, direct: $0.209/s
  • ByteDance Seedance 2.0 at 4K via Runway Dev: $1.50/s

Two things fall out of the normalized list. Direct access to Google's current video model (the Gemini API bills gemini-omni-1.2-flash output at token rates) is cheaper than the same model family resold through at least one aggregator, and the floor of the market has dropped below what any Western flagship charges at its cheapest setting. A 10-second 720p clip costs roughly $0.70 direct from Google, $1.20 on Runway's own model, and $15 on Seedance 2.0 at 4K.

The quality board does not match the price board

Artificial Analysis runs blind human-preference voting on generated video (AA-Video-T2V v2.1 and AA-Video-I2V v2.1). As of this month, its text-to-video leaderboard with audio is led by Alibaba's Wan 3.0 at Elo 1174, followed by Utopai X (1168, no public API), MiniMax H3 at 768p (1159), ByteDance's Dreamina Seedance 2.5 (1156) and FLUX 3 Max (1152). On image-to-video, Seedance 2.5 leads at 1161, with MiniMax H3 fourth at 1134.

The cost implication is awkward for buyers and comfortable for nobody. The top-rated text-to-video model, Wan 3.0, is priced mid-table ($0.10 to $0.20 per second depending on resolution). The cheapest models on the market, MiniMax's H3 family, sit third overall and first among open-weights entries on the same board. And the premium-priced options, Veo 3.1 and 4K tiers, are buying resolution and format guarantees, not preference wins: Veo 3.1 itself is no longer even the headline Google video product on the API, since requests to Omni 1.1 Flash begin routing to Omni 1.2 Flash on December 3, 2026.

The line items the rate card hides

Publishers of per-second pricing attach surcharges that move real budgets:

  • Input video is billed too. Runway's Seedance 2.5 listing charges the output rate plus 50 percent of it per second of input and reference video, with an 80-credit minimum per generation. Hailuo 3 charges 2 credits per reference image on top of the per-second rate.
  • Resolution is a cliff, not a slope. Grok Imagine 1.5 Lite goes from $0.02/s at 480p to $0.14/s at 1080p, a 7x jump. Omni 1.2's own math runs 360p to 4K from $0.023 to $0.209 per second.
  • Professional formats cost extra. ProRes or PNG-sequence output on Runway adds 5 credits per second; HDR profiles add 20 credits per second, 40 above roughly 4 megapixels.
  • Aggregator prices diverge from each other. MiniMax H3 Max text-to-video is $0.03/s on fal and $0.08/s at 768p through Runway's h3_max listing. Same weights, different margin.

The self-hosting footnote

MiniMax published H3 weights on Hugging Face (33B parameters, up to 2K output at 15 seconds with native stereo audio) under the MiniMax H3 Community License Agreement. The model card carries an application form required for users in the US, EU, UK and South Korea, the initial release ships full attention only (sparse attention promised in an update), and the 2K path depends on a hosted regeneration component. "Open weights" here narrows the bill on inference margins; it does not remove the API from the loop for a compliant production deployment.

Bottom line

In late 2026 a second of generated video costs between $0.015 and $1.50 depending on vendor, resolution and how much reference footage you feed the model. The blind-vote quality leader is a mid-priced Chinese model, the price floor is held by Chinese models on aggregators, and the American flagship that anchored every 2025 chart deleted its own API. Teams comparing providers should normalize to dollars per delivered second at their actual resolution, then add input-video billing and format surcharges, because the sticker rate is rarely the invoice.

Sources

video generationPricingAnalysis