Five AI video models for realistic output, with what each one actually costs

Every list of the best AI video models is written by someone who never paid for a generation. This one is not. Every model below was called through the same API on the same day, with the real cost recorded from the response.

Two things became obvious doing it. The gap in quality between the top models is smaller than the marketing suggests. The gap in price is enormous, and it is where the actual decision lies.

How this was tested

All five were called through OpenRouter's video endpoint, which puts every model behind one request format. The same prompt went to each one. Cost came from the usage field in the response, not from a pricing page.

Where a model has a cheaper tier, the cheaper tier was used, because that is what anyone iterating on a prompt would actually run. Prices are per second of output and change often.

1. Veo 3.1

Google's model is the one people mean when they say realistic. Lighting behaves like light, weight shifts correctly through a walk cycle, and rain lands on a jacket rather than passing through it. It reads as filmed rather than rendered.

There are three tiers and the spread between them is brutal. Full Veo 3.1 with audio is $0.40 a second, which is $4.00 for a ten second clip. The Lite tier at 720p without audio is $0.03 a second, which is $0.30 for the same length.

Real cost measured: $0.12 for four seconds at 720p, no audio, on the Lite tier.

The Lite tier is the useful one. It carries most of what makes Veo look right and costs a tenth of the headline price. Move to the full tier only once the prompt is finished.

• Durations: 4, 6 or 8 seconds only
• Resolutions up to 4K on the full tier
• Native audio, which roughly doubles the price
• Slipped out of the arena top ten but still the most brand-safe Western option

2. Seedance 2.5

ByteDance's model is the one to reach for when a sequence has to hold together across shots. It takes reference images, video and audio in a single call, and keeps a character looking like the same character from one shot to the next. That is harder than it sounds and most models fail at it.

It is billed per video token rather than per second, so the cost moves with resolution and length rather than length alone. The 2.5 tier works out around $0.0000107 per token.

Real cost measured: $0.14 for four seconds at 480p on the cheaper 2.0 Mini tier.

Worth knowing that Mini and 2.5 are very different models. Mini is for testing prompts. 2.5 is what the finished thing runs on.

• Durations from 4 up to 30 seconds
• Text to video, image to video with first and last frame control, and reference to video
• Best in class for multi shot work with synced audio
• Four tiers, and the price between them varies by a factor of five

3. Kling 3.0

Kuaishou's model sits between the two above. It is less physically exact than Veo and less consistent across shots than Seedance, but it has the best control stack of the three. First frame and last frame control means the start and end of a shot can be pinned down, which matters enormously when a clip has to cut into an edit.

Standard is $0.084 a second, or $0.126 with audio. Pro is $0.112. Those are flat rates that do not change with resolution, which makes budgeting simple.

• Durations from 3 to 15 seconds
• 16:9, 9:16 and 1:1
• Flat pricing regardless of resolution, unusual in this field
• Good for stylised work, music videos and anything where direction matters more than physics

4. Wan 3.0

Alibaba's model is the value pick. It goes to thirty seconds in a single pass, which none of the others above manage, and 480p costs $0.05 a second. A thirty second clip at 480p is $1.50, which would be roughly $12 on full Veo.

The trade is that it is not as sharp on fine motion. Hands and faces hold up less well under scrutiny. For wide shots, establishing footage and anything where the subject is not the whole frame, the difference is hard to see.

• Durations from 2 to 30 seconds
• 480p, 720p and 1080p, priced at $0.05, $0.10 and $0.20 a second
• Open weights, so it is not locked to one platform
• The cheapest way to generate a lot of footage

5. Sora 2 Pro

OpenAI's model is still the strongest on world consistency. Move a camera through a scene and the walls stay where they were, objects do not morph, perspective does not drift. Glass shattering and water look right in a way the others do not quite manage.

It is also by a distance the most expensive here. $0.30 a second at 720p and $0.50 at 1080p, which puts a twenty second clip at $10. Durations are fixed at 4, 8, 12, 16 or 20 seconds.

This is not a model to iterate on. Get the prompt right somewhere cheap, then spend once.

What the prices actually look like

Ten seconds of video, cheapest sensible setting for each:

• Veo 3.1 Lite, 720p, no audio: $0.30
• Wan 3.0, 480p: $0.50
• Kling 3.0 Standard: $0.84
• Seedance 2.5: varies by resolution, roughly $0.90 to $1.20
• Sora 2 Pro, 720p: $3.00

And the same ten seconds at the top setting of each: Veo 3.1 with audio at 4K is $6.00, Sora 2 Pro at 1080p is $5.00, Wan 3.0 at 1080p is $2.00. The spread between the cheapest and most expensive option in this list is twenty times.

What to actually pick

Footage that has to sit next to real film: Veo 3.1. Nothing else lands lighting and weight as convincingly. Draft on Lite, finish on the full tier.

A sequence with a recurring character: Seedance 2.5. The reference inputs and shot to shot consistency are what it is built for.

Anything cutting into a timeline: Kling 3.0. First and last frame control makes the clip fit rather than approximately fit.

A lot of footage on a small budget: Wan 3.0 at 480p. Thirty seconds for $1.50 is not close to matchable.

One hero shot where physics has to survive inspection: Sora 2 Pro. Once. After the prompt is settled elsewhere.

The part nobody mentions

Prompt iteration is where the money goes, not the final render. Ten attempts at four seconds on a cheap tier costs about a dollar. Ten attempts at ten seconds on full Veo with audio costs forty.

So the workflow that actually saves money is boring. Write the prompt. Test it at the shortest duration and lowest resolution the model allows. Keep testing until the composition, the motion and the framing are all right. Only then raise the settings and spend properly, once.

None of these models are free and none of them will be. The video subscriptions that feel free are subsidised by people who generate five clips a month. On an API you pay what it costs to run.