Question 1

What is an Elo score in AI video?

Accepted Answer

Elo is a relative rating from head-to-head matchups. We run two models on the identical prompt, hide the vendor names, and record which render human raters prefer. Winning against a strong model raises a score more than winning against a weak one, so 1687 vs 1520 means the first model wins that pairing roughly 72% of the time.

Question 2

How do you test the models?

Accepted Answer

Every model gets the same prompt set, the same seed budget and the same resolution ceiling its API supports. Renders are stripped of watermarks and vendor metadata, then scored blind on eight axes: prompt adherence, motion realism, character consistency, temporal coherence, lipsync accuracy, native audio quality, speed to first frame and cost efficiency. We re-run the full suite weekly.

Question 3

Which model is best for lipsync?

Accepted Answer

Veo 3.1 currently leads lipsync accuracy at 88 out of 100, with Seedance 2.0 close behind at 85 and cheaper per second. If you need dialogue-driven scenes with native audio in one pass, those two are the only safe defaults today.

Question 4

Which model is cheapest per second?

Accepted Answer

Cheapest list price is rarely cheapest in practice, because a model that needs four retries costs four times its sticker price. We publish price per second per provider on the pricing dashboard alongside Elo, so you can compare cost per video that actually works rather than cost per attempt.

Question 5

Is my content commercial-safe?

Accepted Answer

It depends on the provider licence, not the benchmark. Closed-weight models from ByteDance, Google DeepMind and Runway grant commercial use through their paid tiers but restrict likeness and trademark content. Open-weight models let you self-host but shift indemnity onto you. Always read the provider terms before shipping client work, and watch our deprecation notices so a sunset does not break a live pipeline.