RANKING · BUDGET TIER
The cheapest AI video generators, ranked
| Model | $ / second | 10s clip | 1 min | Audio | Max clip |
|---|---|---|---|---|---|
| Vidu 2.0Shengshu | $0.04 | $0.40 | $2.40 | — | 10s |
| Veo 3.1 LiteGoogle | $0.05 | $0.50 | $3.00 | Yes | 8s |
| Gen-4 TurboRunway | $0.05 | $0.50 | $3.00 | — | 10s |
| Wan 2.6Alibaba | $0.05 | $0.50 | $3.00 | — | 10s |
| Sora 2OpenAI | $0.10 | $1.00 | $6.00 | Yes | 12s |
| Kling 3.0Kuaishou | $0.10 | $1.00 | $6.00 | — | 120s |
| Seedance 2.0ByteDance | $0.14 | $1.40 | $8.40 | — | 10s |
| Veo 3.1 FastGoogle | $0.15 | $1.50 | $9.00 | Yes | 8s |
| Gen-4.5Runway | $0.25 | $2.50 | $15.00 | — | 10s |
| Sora 2 ProOpenAI | $0.30 | $3.00 | $18.00 | Yes | 25s |
| Veo 3.1Google | $0.40 | $4.00 | $24.00 | Yes | 8s |
Rates shown at each model's base tier, verified July 2026 from official vendor pricing pages and documentation. Vendors change prices without notice — see methodology.
The cheapest tiers in 2026
On verified per-second pricing, the budget podium is: Vidu 2.0 (~$0.04/s), Wan 2.6 and Runway Gen-4 Turbo ($0.05/s), and Veo 3.1 Lite (~$0.05/s) — with Kling's consumer subscription an honorable mention at an effective ~$0.04/s in standard mode. Seedance 2.0's 480p tier ($0.065/s) is the sleeper pick for drafting.
Why "cheapest" is a trap question
The real cost of a usable clip is price × attempts. Budget models typically have weaker prompt adherence, so a $0.05/s model needing four takes costs more than a $0.15/s model that lands in one. Cheap tiers shine when failure is cheap to detect — B-roll, ambient shots, loops — and premium tiers earn their price on anything with faces, hands or text in frame.
The genuinely free routes
- Google Cloud's $300 new-account credit covers roughly 100 minutes of Veo 3.1 Lite — the largest free allowance in the market.
- Kling's free monthly credits refresh and cover casual social use.
- Self-hosting Wan (open-source) makes marginal cost near zero if you own the GPU.
Run your own numbers. The cost calculator applies your clip length, resolution and a realistic retake buffer across every model at once.