A comparative test of 12 video generation models was conducted on October 9, 2026, using a single prompt to ensure consistency. The task required an 8-second, 720p video featuring a street food vendor, specific lighting, and physics-based elements like fire and steam. All generations were processed through the SYNTX.AI platform to maintain uniform accounting.
Luma Ray 3.14 topped the ranking with a score of 9.1 out of 10, costing 5 tokens. While it is the most affordable option, it lacks native audio and character reference features. Seedance 2.5 delivered the highest visual quality but was the most expensive, consuming 138.7 tokens and requiring six minutes of generation time.
Gemini Omni Flash 1.1 proved to be a cost-effective solution, ranking second in efficiency with 10.5 tokens per generation. Kling 3.0 performed well for multi-shot sequences, offering audio and lip-sync capabilities. Wan 2.7 ranked lowest in the test, scoring 3.2 out of 10 and requiring ten minutes for a single generation.
The evaluation criteria weighted visual quality and prompt adherence at 80%, with speed and token efficiency accounting for 10% each. The test highlighted that model performance can vary significantly based on the assigned weighting of these metrics, with Seedance 2.5 being particularly sensitive to changes in evaluation parameters.
Other notable models included FLUX 3 Video, which balanced quality and cost, and Grok Imagine 1.5 Preview, which requires a starting image. Veo 3.1, despite its 4K capabilities, ranked seventh due to a single-pass limitation and high resource consumption. MiniMax H3 (Hailuo 3.0) and Seedance 2.0 were also evaluated for their specific handling of physics and multi-language support.
