Three questions read any new AI video generation model in minutes. The first asks how the model holds an object across frames. The second asks what it compresses a clip into, and the third asks how long one clip runs and what it costs.
The answers change with every release, and the questions do not. They follow from how video generation works. A model must keep an object consistent over time, work on a clip small enough to afford, and ship at a price you can plan around.
TLDR
Test object memory yourself. Pull 4 frames from one clip and compare color, shape and size.
Look up the latent size in the technical report. A smaller latent costs less to denoise.
Multiply the price per second by the clip length and by your tries.
Treat every launch-day fact as dated, and re-check it before you plan on it.
Realistic video no longer proves that an event happened.
Three Questions Read Any New Video Model Release
Run one release two ways.
Read the launch thread, and you see the best clip out of many tries, with a price you must guess.
Ask the three questions, and you pull four frames from a test clip, read the latent size, and multiply the price by your tries.
The second path ends in a go or no go before you spend a credit.
Question 1. A Four-Frame Test Shows If It Holds an Object
A video is a stack of frames, and each frame is a fresh guess unless the model can look back. Early systems drew each frame alone. Objects changed shape and lighting shifted between frames.
Two fixes exist, and a release page usually names one of them. One adds temporal layers to an image model, as in the video latent diffusion paper. The other turns the clip into spacetime patches and runs a transformer over them, as in OpenAI's technical report on video generation.
A transformer scores every token against the others, a step called attention. That lets the model check the cube in frame 4 against the cube in frame 1. OpenAI's report states that model details are not included, so a vendor's claim tells you less than a test does.
Run the test with one plain object in the prompt. Export a frame every second, then compare three things:
Color. Does the object keep its color?
Shape. Do the edges stay straight?
Size. Does it grow or shrink while the camera holds still?
Here is that test on one real clip.
Example. On 2026-10-13 I gave MiniMax Hailuo 2.3 this 6-second prompt at 768p: "A single red cube rolls slowly across a wooden table from left to right, one full turn, steady side-on camera, soft daylight, plain white wall behind it." This was one run. The cube stayed red in frames at 0, 1, 3 and 5 seconds, and its apparent size shifted as it turned. It tumbled with small hops instead of rolling, and one clip proves nothing about the next.
Question 2. Compression Decides Why One Clip Costs Less to Make
Every denoising step touches every number in the clip, so the count of numbers sets the cost. Raw pixels scale badly. A 16-frame clip at 64 by 64 pixels already holds 196,608 values, and an HD frame is far larger.
A latent model fixes that in three moves. An encoder compresses the clip, the model denoises the small version, and a decoder turns the result back into pixels. The latent diffusion paper reports a large compute saving over working on pixels.
Take a common image autoencoder as a worked example. The CompVis repository lists a downsampling factor of 8, and the model uses 4 latent channels. The same clip drops from 196,608 values to 4,096, which is 48 times fewer.
Rule of thumb. Video models also compress time, as OpenAI's report says, so the real ratio is larger than this image-only example.
Open the technical report and search for "latent". Write down the compression factor in space and in time. A model with no report leaves question 2 unanswered, and you price it by trial.
Question 3. Clip Length Times Price Times Tries Is Your Real Cost
A clip has a length and a price per second. Both sit on the vendor's pricing page, and both change. Then multiply by a third number, the tries it takes to get one usable clip.
Seconds per clip. The model page states it.
Price per second. The pricing page states it.
Tries per usable clip. Only your own test shows it.
Tries vary widely. One creator on X says a good clip takes 3 to 4 tries, and he mocks a report of 1 usable clip in 87. Both numbers are self-reported, but the gap shows that your yield sets your cost.
One creator's post shows how far yield can swing.
Example. On 2026-10-13, Google's Gemini API pricing page lists Veo 3.1 Standard at $0.40 per second at 1080p and $0.60 at 4K, both with audio. An 8-second 4K clip costs $4.80, and 4 tries cost $19.20. The Fast tier at 4K costs $0.30 per second, or $2.40 per clip.
Tries are the cost you control, and a structured prompt is where to start. I built AI Cinema Director, a skill that writes shot prompts with camera, lens and timing for a video model. It writes text only, so you generate the video in Higgsfield, which bills its own credits.
The same multiply-by-tries habit prices token bills, as in part 1 of the AI cost series.
Realistic Video Ended Proof by Footage
A realistic clip no longer shows that an event happened. Researchers tested this by prompting a video model with false claims, and it produced convincing video for most of them. NewsGuard also found that the watermark marking each clip as Sora's came off easily.
OpenAI introduced Sora 2 with this launch video.
Watch out. NewsGuard prompted Sora 2 with 20 false claims on October 14, 2025. It got video for 16 of them, 11 on the first try, including all 5 claims tied to Russian disinformation. OpenAI closed the Sora app on April 26, 2026 and its API on September 24, 2026, so the product is gone and the finding stands.
NewsGuard, a misinformation research firm, published the full test in its newsletter.
Treat video as a claim until you have the source file, the uploading account and a second angle. The trust problem outlasts every model, and it will shape the next decade of visual media more than resolution does.
Four Questions a Newcomer Asks About AI Video Models
Do I need the technical report for every model? Only for question 2. Question 1 needs one test clip, and question 3 needs the pricing page.
How many frames should I compare? Four, one second apart, from one clip. Repeat on three prompts before you trust the result.
Why not trust the demo reel? A reel shows chosen clips, and it hides how many tries each one took.
Can code answer these for me? Question 3 is arithmetic you can script. For a judgment call on a written release page, ask a model that returns a typed answer, such as Jev.
The Appendix Holds the Commands and the Formulas
Question 1, export one frame per second from a test clip:
ffmpeg -i clip.mp4 -vf fps=1 frame_%02d.pngQuestion 2, count the values before and after compression:
raw = frames x height x width x 3
16 x 64 x 64 x 3 = 196,608
latent = frames x (height / 8) x (width / 8) x 4
16 x 8 x 8 x 4 = 4,096Question 3, price one usable clip:
cost = seconds per clip x price per second x tries
= 8 x 0.60 x 4 = 19.20
Man, perfect timing. Was litteraly just thinking about deepfakes in nature docs.
The technical progression is explained cleanly. The trust problem now matters more than quality.