The Same 15 Seconds Costs $1.05, $1.50, or $3.47

For most of the year, comparing video generation models meant comparing company demos. Google, ByteDance and Alibaba each published a single-shot clip that flattered their own model. That changed on November 6, when Google released Gemini Omni 1.2 Flash and the first independent test of it appeared alongside Artificial Analysis numbers and side-by-side runs from other creators.
The most useful thing in the pile is the price table. A 15-second clip at 720p costs $1.05 on Omni 1.2, $1.50 on Wan 3.0 and about $3.47 on Seedance 2.5. At 1080p the same clip runs $1.56, $3.00 and $8.54. Alibaba also runs a roughly 30 percent promotion on Wan 3.0, which puts 1080p as low as $0.14 per second.
The spread runs three-fold on the cheapest tier and five-fold at 1080p, and the cheapest model leads few of the categories.
Where each model actually wins
Artificial Analysis, as of November 6, puts Seedance 2.5 first in 13 of 20 categories, including dialogue and lip sync at a score of 1213. Image-to-video is its strongest mode. Gemini Omni 1.1 ranked twentieth of the twenty in human anatomy and physics.
Editing is where Wan 3.0 leads, taking complex edits, object edits and visual effects, while Seedance takes sound and speech editing.
The same-prompt tests from outside the companies are more revealing than the leaderboard. Krea ran Seedance 2.5 and Wan 3.0 at 720p with identical seeds and its prompt rewriting turned off. In a skateboard shot, Seedance kept the skater on the path described in the prompt while Wan sent him toward the camera. In a 20-second single take following a cyclist, Seedance held the shot to the end and Wan cut away at 13.6 seconds. Wan also rendered its eight-second clips roughly twice as fast. YouTube creator JSFILMZ ran both models side by side on matched prompts and posted the results.
Omni 1.2 has almost no independent testing yet. The closest is Segmind, which ran twenty tests on launch day and called Omni 1.2 Flash the most instruction-exact video model it had tested, good for explainers and animated text. Segmind also found that photorealistic people looked floaty, faces fell short of Seedance, and roughly one instruction in four was skipped.
Google itself admits there are small changes from the first frame to the last, and lists fast sports action and complex physics as work in progress. Brand logos in source footage can trigger a copyright block. Every clip carries an invisible SynthID watermark.
What the request numbers say
OpenRouter's video rankings, data through November 6, show HeyGen Video first with 65,000 weekly requests, Veo 3.1 Lite second at 50,000, Seedance 2.5 third at 35,000, Wan 3.0 fourth at 31,000, and mischief 1 ninth after a 481 percent jump.
The market share by author is the number worth sitting with. ByteDance holds 26.1 percent of video requests, HeyGen 19.7 percent, Google 18.5 percent, SpaceXAI 12.2 percent, Alibaba 9.3 percent after a 53 percent drop, Kwaivgi 6.0 percent and MiniMax 5.0 percent.

The most-requested video model on the largest aggregator is a digital human. Nobody's product launch is winning the volume; talking-head explainers and training videos are. Seedance and Wan lead the cinematic category, and their share of total requests is a rounding error next to the avatar traffic.
The demand that is not cinematic
Look again at where the requests go. HeyGen Video leads the table with 65,000 weekly requests, and HeyGen is an avatar company. Enterprise security and enablement teams have started replacing recurring live product walkthroughs with AI-generated presenter videos, which by their own accounting saves more than 100 hours a year per team. Marketer adoption sits around 63 percent according to a 2026 survey cited by Colossyan.
Meanwhile native audio has moved from feature to default across the category. Veo 3.1 Audio and its fast variant generate sound matched to the picture, Kling's audio tiers produce stereo with improved lip sync, and Canva's text-to-video tool produces clips with synchronized dialogue and effects.
The split this creates is useful. Avatar and talking-head work wants the character to stay consistent, the voice to stay steady and the delivery to land. Cinematic work wants the shot to impress for eight seconds. These are different requirements and different models serve them, which is why a single leaderboard covering both tells you less than the request split does.
What changed when Sora left
The exit of Sora from the web and app in April and from the API in September did more than remove a competitor. It reset what people expect to pay for and what they expect to be available, and it left the remaining platforms competing on the axes that actually determine usage: how long a clip can be, whether the audio arrives with it, and what happens to the character between shots.
Two years ago the reasonable default for a new product was to build on OpenAI's video API. That default is gone. Anything built on it needs a replacement, and the replacements now differ enough in price and length that the choice of model is a business decision rather than a technical one.
The reseller layer adds one more wrinkle worth pricing in. Official API rates are the floor, not the street price. fal, for example, lists Seedance 2.5 at roughly twice the BytePlus rate. Teams who test on a marketplace and then plan production budgets against that number are in for a surprise, and the gap between official and resale pricing is wide enough to change which model a project picks.
Choosing with these numbers
If the job is dialogue, lip sync or audio, Seedance 2.5 is the safer bet at roughly three times the price. If the job is editing an existing clip, Wan 3.0 leads and its promotional pricing makes the cheapest per-second cost on the board. If the job is following instructions literally in an explainer, Omni 1.2 Flash is the one early testers keep praising, and it is the cheapest of the three at 720p, though Segmind's own tests took 40 to 70 seconds per clip.
The right comparison is price per usable clip. A model that needs four attempts to hold a face through a 15-second take costs four times its sticker price, and the leaderboard cannot see that.
None of the three companies has published a blind, same-prompt, same-seed comparison across all three models. Until somebody runs it, treat every cross-model quality claim in this category, including the ones in this article, as directional.
One thing is clear from the request data. Cheaper models are only cheap if the output survives review, and review is where the cost lives for most teams shipping video at volume.
Related articles
METR Spent 11 Weeks Inside Anthropic. The Number Is 1.9, and Most of the Evidence Is Blacked Out.
A policy threshold with a number in it, an auditor who says it may have been crossed, and evidence nobody outside can check.
One Percent of Firms Pay 80% of the AI Bill. That Cohort Just Stopped Growing.
A $600 million revenue cut, a 26.8% price decline, and a cohort that slipped below $10,000 per employee.
Claude Can Now Read Your Bank Statements. Anthropic Deliberately Gave It No Hands.
The read-only design is the product, and unlinking does not retract what is already in a chat.
Adobe Stopped Asking Marketers to Open Adobe
For a decade the strategy was own the interface. Adobe just bet the opposite.