Everyone reviews the press release. We run the models.
AI video model benchmarks
and leaderboard
The same 40 prompts through every AI video model that matters: hands, on-screen text, liquids, multi-shot continuity, lip-sync. We publish every clip, every score and where each model is cheapest to run.
The same 40 prompts through every AI video model. Every clip public, and where each one is cheapest to run.
See the leaderboard Price radar
- Models scored
- 16
- Clips judged
- 698
- Prices tracked
- 14,075
- Price check
- 2× a day
Same prompt, every model
360 orbit: The hardest camera move. Requires the model to hold a whole scene in 3D.
“A full 360-degree orbit around a seated woman reading a book, background geometry stays consistent”
Leaderboard
Full board →- 1
Kling 3.0 Turbo
4.84 - 2
FLUX 3
4.84 - 3
Seedance 2.0
4.84 - 4
Google Gemini Omni
4.82 - 5
MiniMax H3
4.79 - 6
Seedance 2.5
4.79
- 7
Vivix W1
4.77 - 8
OpenAI Sora 2
4.76
Overall score out of 5 across 40 prompts, judged blind. Bars start at 3.0 so the gaps are visible.
Where it’s cheapest
Price radar →Price of one clip, same spec on every platform, checked 6h ago.
-
Kling 3.0 Turbo
720p · 5s · audio
≈$0.26
Higgsfield up to $0.56 on 4 others
-
FLUX 3
720p · 5s · audio
$0.75
Black Forest Labs API up to $0.85 on 1 other
-
Seedance 2.0
720p · 5s
$0.73
Lumeta up to $1.80 on 9 others
-
Google Gemini Omni
720p · 5s
$0.51
Google Gemini API up to $1.06 on 3 others
-
MiniMax H3
768p · 5s
$0.20
Lumeta up to $0.68 on 6 others
How we test
The same prompts, every time
Chosen because they break models: fingers crossing, legible on-screen text, a horse’s gait, one character across three shots.
Clips, all public
Every score links to the video it came from. Disagree with a number? Watch the evidence.
Scored without the model’s name
A vision model judges each clip against its prompt. Any failure it reports is checked against real frames before it counts.
Prices, normalized
Credits, per-second rates and plans turned into USD for one clip, so a platform’s price means the same as the next one.
Latest
All news →Google turns off the three Veo 3.1 preview endpoints today, with Gemini Omni 1.1 Flash as the only named replacement
Developers calling Veo 3.1 through Google's consumer-facing Gemini API hit their deprecation date today. The migration target is a different model with a different call shape, and it is the second forced Veo move in four months.
22 OctNVIDIA ships a real-time surgical video model and states that visual realism is not proof of accuracy
A video generation release that declines to claim its own output is correct. NVIDIA's new surgical world model streams interactively on one GPU, and the announcement attaches no frame rate, no clinical evaluation and no comparison against a conventional simulator.
21 OctVideoGen unveils Model 1 with a ten-minute ceiling and no published price, benchmark or third-party number
A Y Combinator-backed company put a ten-minute number and a 50% cost reduction in the same press release, then listed pricing, benchmarks and outside validation as things that have not happened yet.
21 OctDecisionPivot scores 40 automated judges across 40,033 items, and 19 of them land under a guessing baseline
The benchmark is built to show what aggregate accuracy hides in the small classifier models now sold as automated judges. Its 40,033 items are entirely text, which is the gap for anyone scoring generated video or images.