October 2026 Video Model Rankings: Luma Ray 3.14 Wins on a Single Prompt for 5 Tokens

Twelve video models, one task, 8 seconds of footage

Twelve neural networks for video, one text prompt, identical conditions: 720p, 16:9, 8 seconds, with no hints about each model's quirks. The test launched on October 9, 2026. By total score, Luma Ray 3.14 won (9.1 out of 10), completing the task in two minutes for 5 tokens — the cheapest in the entire lineup. Seedance 2.5 completed the task with no shortcuts and delivered the best picture, but it cost 138.7 tokens — 27 times more than the leader.

In March 2026, OpenAI shut down Sora, the model everyone compared everything to just a year earlier. No single product fully took its place in the rankings: twelve generators split the spot instead. The spec sheets for the rest look nearly identical (1080p, sound, 10–15 seconds, support for reference images), and typical rankings only compare those numbers. The real difference in how a model performs only shows up on an actual task.

Methodology: why 8 seconds and 720p

720p was chosen deliberately, even though half the models can work at higher resolution. Comparing 4K against 720p would be pointless: the resolution gap would swamp any quality difference. All twelve were shot under identical conditions.

8 seconds was chosen because Veo 3.1 no longer does more than that in a single pass. Pitting 8 seconds against 15 wouldn't be fair. Each model was run only once (in practice, multiple attempts are usually needed). It's more interesting to see whether models succeed on the first try. Even in this scenario, Veo 3.1 delivered one of the weakest results in the top half.

The test checked: whether objects stay consistent from the first frame to the last, how fire, steam, smoke, and water behave, and whether the model actually followed the prompt. Sound was scored separately: Runway and Luma don't generate their own sound. Grok Imagine 1.5 Preview and Runway Gen-4 Turbo only work from a starting frame rather than text. For those two, a starting frame was generated using their own image models (Grok Frames and Runway Frames), and video was then generated from that image — still a single run each, same as everyone else.

Results: twelve models, ranked

Luma Ray 3.14 (9.1/10, 5 tokens, 2 minutes) — the best price-to-speed ratio. Requires color correction, doesn't generate sound, no character-reference support.

Seedance 2.5 (138.7 tokens, 6 minutes) — the only model to complete the task with zero shortcuts, best image quality, the most expensive in the test.

Gemini Omni Flash 1.1 (10.5 tokens, 1 minute) — economical and fast, with a 720p ceiling and a ten-second limit.

Kling 3.0 (3–15 seconds) — flawless sound and lip sync, handles multi-shot sequences and dialogue well.

Wan 2.7 (3.2/10, 10 minutes of generation) — loses across all three scoring categories.

Veo 3.1 — the single run turned out to be the weakest result in the upper half of the list.

Grok Imagine 1.5 Preview — only works from a starting frame; can't be run from text alone.

Runway Gen-4 Turbo — only works from a starting frame.

Four more models from the list appear in the summary table, but the source material doesn't break them down individually.

How the scores were calculated

The final score is made up of three parts: image quality and prompt adherence (80%), generation speed (10%), and token efficiency (10%). Quality carries the most weight because people commission a clip for the picture. Speed and price only adjust that baseline.

The speed scale is linear: 1 minute = 10 points, 10 minutes = 1 point. Token efficiency: 5 tokens = 10 points, 138.7 tokens = 1 point (tokens are scored by the spread across the list, not by the raw number).

All twelve runs went through one environment, one account, and one token-accounting format. Every service counts tokens differently, and usage can't be added up across different platforms. Here, a single unit of measurement applies to the whole test, which is why the numbers are comparable.

The test prompt was deliberately complex (a street, noodles, fire, steam, rain, neon, camera movement, close-up hands, 35mm film, shallow depth of field) — a boring task wouldn't have revealed the differences.