Price is per image, per generation or per second of video. Check the unit before you compare two rows. Ten seconds at $0.10 a second costs $1, not 10 cents.
Takes is what you give the model to start: a text prompt, a still image, a video whose movement it copies, or audio it lip-syncs to.
LoRA means you can load your own trained character fileLoRA A LoRA is a small file trained on images of one subject. It teaches a general image model to reproduce that specific face or style. Open the glossary. It is the classic way to keep one face the same.
Open weights means you can download the model and run it yourself. It does not mean you are allowed to use it commercially.

Video models billed per second

Sorted cheapest first. The price is for one second of finished video at the default resolution. The grey line under it shows the same rate for a 10-second clip.

Model	Maker	Price	Takes	LoRA	Notes
Wan 2.2 Ultra Fastopen weights	Alibaba	$0.01 per second≈ $0.10 for a 10 s clip	image-to-video	—	Priced per second and not per clip. You get a hundred seconds per dollar, but the model is one generation behind.
InfiniteTalk	MeiGen-AI	$0.03 per second≈ $0.30 for a 10 s clip	avatar-lipsync	—	Does lip-sync, not animation. That is a different job from making a still image move, and you only need it if your persona speaks.
Wan 2.2 Animateopen weights	Alibaba	$0.04 per second≈ $0.40 for a 10 s clip	motion-transfer	—	Copies existing motion onto a character and does not invent it. That keeps the movement believable.
MiniMax H3open weights	MiniMax	$0.04 per second≈ $0.40 for a 10 s clip	image-to-video, text-to-video, video-extend	—	The open-weights edition, which WaveSpeed runs on its own servers. A 5-second clip at the default 480p lists at $0.20. On 5 October 2026 WaveSpeed showed it at 50% off, so that clip cost $0.10. MiniMax’s official H3 on WaveSpeed is a separate model at $0.10 per second (768p).
MiniMax H3 (LoRA)open weights	MiniMax	$0.05 per second≈ $0.50 for a 10 s clip	image-to-video, reference-to-video	Yes	Takes up to three LoRA files, but only LoRAs trained for MiniMax H3. WaveSpeed says a LoRA must match the base model exactly, and on 5 October 2026 it had no H3 trainer. Reference-to-video is $0.06 per second. 50% off on 5 October: $0.125 for a 5-second image-to-video clip.
Kling 3.0 Std	Kuaishou	$0.084 per second≈ $0.84 for a 10 s clip	image-to-video, text-to-video	—	The standard tier of the Kling line, which has long been a default choice for character video. Price without sound, the default. With sound it costs 1.5 times as much.
Gemini Omni 1.1 Flash	Google	$0.1 per second≈ $1.00 for a 10 s clip	text-to-video, reference-to-video	—	Google’s video model with sound, which also takes up to ten reference images. $0.10 per second at the default 720p on WaveSpeed. Google’s own Gemini API states about the same: roughly $0.10 per second of 720p video. 360p costs $0.03 per second.
Wan 3.0	Alibaba	$0.1 per second≈ $1.00 for a 10 s clip	image-to-video, reference-to-video, text-to-video	—	Takes 2 to 30 seconds in one generation, and several kinds of reference (multimodal), not only a first frame. 480p costs $0.05 per second. On 5 October 2026 WaveSpeed took 5% off.
Kling 3.0 Std Motion Control	Kuaishou	$0.126 per second≈ $1.26 for a 10 s clip	motion-transfer	—	Takes a character image and a driving video. The character then moves the way the person in the video moved. Billed $0.63 per 5 seconds, up to 30 seconds.
Veo 3.1 Fast	Google	$0.15 per second≈ $1.50 for a 10 s clip	text-to-video, image-to-video	—	Price with sound, which is on by default; without sound it is $0.10 per second. Google’s own Gemini API charged $0.10 per second at 720p with sound until it shut the Veo 3.1 preview down on 22 October 2026. Google Cloud’s Agent Platform lists the same $0.10 for the GA model (21 October 2026). Its text-to-video results are the ones people show off most often.
Wan 3.0 Prime	Alibaba	$0.15 per second≈ $1.50 for a 10 s clip	image-to-video, reference-to-video, text-to-video	—	A faster variant of Wan 3.0 at 1.5 times the price. On 5 October 2026 WaveSpeed took 5% off.
Kling 4.0	Kuaishou	$0.157 per second≈ $1.57 for a 10 s clip	text-to-video, image-to-video, reference-to-video, video-edit	—	Price on Kling’s own API at 720p without an input video, read in early access on 21 October 2026. Kling’s prepaid API packages did not cover 4.0 yet. 1080p costs $0.235 per second, and there is no 4K price. With an input video it is $0.094 per second, but Kling bills at least twice the clip length. Makes 3 to 30 seconds in one generation. Kling does not say whether sound costs extra.
Seedance 2.0 Fast	ByteDance	$0.2 per second≈ $2.00 for a 10 s clip	image-to-video, text-to-video	—	At the default 720p a ten-second clip costs $2, not 20 cents. At 480p it is $0.10 per second. Check the unit before you compare.
Seedance 2.5	ByteDance	$0.36 per second≈ $3.60 for a 10 s clip	image-to-video, text-to-video, video-edit, video-extend	—	The premium end. A 5-second clip at the default 720p lists at $1.80; 480p costs $0.18 per second. On 5 October 2026 WaveSpeed took 10% off, so that clip cost $1.62. With the video-edit and video-extend modes you can repair a clip that is almost right and do not have to generate it again.

Wan 2.2 Ultra Fastopen weights
$0.01 per second≈ $0.10 for a 10 s clip
Maker
Alibaba
Takes
image-to-video
Custom LoRA
No
Priced per second and not per clip. You get a hundred seconds per dollar, but the model is one generation behind.
InfiniteTalk
$0.03 per second≈ $0.30 for a 10 s clip
Maker
MeiGen-AI
Takes
avatar-lipsync
Custom LoRA
No
Does lip-sync, not animation. That is a different job from making a still image move, and you only need it if your persona speaks.
Wan 2.2 Animateopen weights
$0.04 per second≈ $0.40 for a 10 s clip
Maker
Alibaba
Takes
motion-transfer
Custom LoRA
No
Copies existing motion onto a character and does not invent it. That keeps the movement believable.
MiniMax H3open weights
$0.04 per second≈ $0.40 for a 10 s clip
Maker
MiniMax
Takes
image-to-video, text-to-video, video-extend
Custom LoRA
No
The open-weights edition, which WaveSpeed runs on its own servers. A 5-second clip at the default 480p lists at $0.20. On 5 October 2026 WaveSpeed showed it at 50% off, so that clip cost $0.10. MiniMax’s official H3 on WaveSpeed is a separate model at $0.10 per second (768p).
MiniMax H3 (LoRA)open weights
$0.05 per second≈ $0.50 for a 10 s clip
Maker
MiniMax
Takes
image-to-video, reference-to-video
Custom LoRA
Yes
Takes up to three LoRA files, but only LoRAs trained for MiniMax H3. WaveSpeed says a LoRA must match the base model exactly, and on 5 October 2026 it had no H3 trainer. Reference-to-video is $0.06 per second. 50% off on 5 October: $0.125 for a 5-second image-to-video clip.
Kling 3.0 Std
$0.084 per second≈ $0.84 for a 10 s clip
Maker
Kuaishou
Takes
image-to-video, text-to-video
Custom LoRA
No
The standard tier of the Kling line, which has long been a default choice for character video. Price without sound, the default. With sound it costs 1.5 times as much.
Gemini Omni 1.1 Flash
$0.1 per second≈ $1.00 for a 10 s clip
Maker
Google
Takes
text-to-video, reference-to-video
Custom LoRA
No
Google’s video model with sound, which also takes up to ten reference images. $0.10 per second at the default 720p on WaveSpeed. Google’s own Gemini API states about the same: roughly $0.10 per second of 720p video. 360p costs $0.03 per second.
Wan 3.0
$0.1 per second≈ $1.00 for a 10 s clip
Maker
Alibaba
Takes
image-to-video, reference-to-video, text-to-video
Custom LoRA
No
Takes 2 to 30 seconds in one generation, and several kinds of reference (multimodal), not only a first frame. 480p costs $0.05 per second. On 5 October 2026 WaveSpeed took 5% off.
Kling 3.0 Std Motion Control
$0.126 per second≈ $1.26 for a 10 s clip
Maker
Kuaishou
Takes
motion-transfer
Custom LoRA
No
Takes a character image and a driving video. The character then moves the way the person in the video moved. Billed $0.63 per 5 seconds, up to 30 seconds.
Veo 3.1 Fast
$0.15 per second≈ $1.50 for a 10 s clip
Maker
Google
Takes
text-to-video, image-to-video
Custom LoRA
No
Price with sound, which is on by default; without sound it is $0.10 per second. Google’s own Gemini API charged $0.10 per second at 720p with sound until it shut the Veo 3.1 preview down on 22 October 2026. Google Cloud’s Agent Platform lists the same $0.10 for the GA model (21 October 2026). Its text-to-video results are the ones people show off most often.
Wan 3.0 Prime
$0.15 per second≈ $1.50 for a 10 s clip
Maker
Alibaba
Takes
image-to-video, reference-to-video, text-to-video
Custom LoRA
No
A faster variant of Wan 3.0 at 1.5 times the price. On 5 October 2026 WaveSpeed took 5% off.
Kling 4.0
$0.157 per second≈ $1.57 for a 10 s clip
Maker
Kuaishou
Takes
text-to-video, image-to-video, reference-to-video, video-edit
Custom LoRA
No
Price on Kling’s own API at 720p without an input video, read in early access on 21 October 2026. Kling’s prepaid API packages did not cover 4.0 yet. 1080p costs $0.235 per second, and there is no 4K price. With an input video it is $0.094 per second, but Kling bills at least twice the clip length. Makes 3 to 30 seconds in one generation. Kling does not say whether sound costs extra.
Seedance 2.0 Fast
$0.2 per second≈ $2.00 for a 10 s clip
Maker
ByteDance
Takes
image-to-video, text-to-video
Custom LoRA
No
At the default 720p a ten-second clip costs $2, not 20 cents. At 480p it is $0.10 per second. Check the unit before you compare.
Seedance 2.5
$0.36 per second≈ $3.60 for a 10 s clip
Maker
ByteDance
Takes
image-to-video, text-to-video, video-edit, video-extend
Custom LoRA
No
The premium end. A 5-second clip at the default 720p lists at $1.80; 480p costs $0.18 per second. On 5 October 2026 WaveSpeed took 10% off, so that clip cost $1.62. With the video-edit and video-extend modes you can repair a clip that is almost right and do not have to generate it again.

Image models

Sorted cheapest first. The most expensive model costs roughly thirty times as much as the cheapest. The price decides how many images you can afford to throw away.

Model	Maker	Price	Takes	LoRA	Notes
Z-Image-Turbo	WaveSpeed	$0.005 per image	text-to-image	—	The cheapest image model on this list. You get two hundred images per dollar, so you can generate a lot and throw most of it away.
Flux 2 Klein	Black Forest Labs	$0.008 per image	text-to-image	—	Most persona work has been built on the Flux family. Klein is the cheap tier of the current generation.
GPT Image 2.5 Edit	OpenAI	$0.039 per generation	image-edit	—	OpenAI's newest image model (September 2026), also in ChatGPT. Very realistic from reference images of your own persona. Price for 1K at medium quality with one reference image; each extra reference adds about $0.015.
Seedream 4.5	ByteDance	$0.04 per image	text-to-image	—	ByteDance’s image model. Its price sits between the cheap bulk models and the premium ones.
Seedream V5.0 Pro Edit	ByteDance	$0.045 per generation	image-edit	—	Edits one or more images you give it and does not generate from nothing. This is the mode you need to fix a single frame. On 5 October 2026 WaveSpeed took 10% off, so one edit cost $0.0405.
GPT Image 2 Edit	OpenAI	$0.07 per generation	image-edit	—	Edits an image from instructions in normal language. Good at text inside an image, which most models still get wrong. On 5 October 2026 WaveSpeed took 5% off, so one edit cost $0.0665.
Nano Banana 2	Google	$0.07 per image	text-to-image, image-edit	—	The standard tier of Google’s image models. Price at 1K. On 5 October 2026 WaveSpeed took 10% off, so one image cost $0.063.
Nano Banana Pro	Google	$0.14 per image	text-to-image, image-edit	—	The premium tier. Most subscription credit systems base their prices on this model. On 5 October 2026 WaveSpeed took 10% off, so one image cost $0.126.

Z-Image-Turbo
$0.005 per image
Maker
WaveSpeed
Takes
text-to-image
Custom LoRA
No
The cheapest image model on this list. You get two hundred images per dollar, so you can generate a lot and throw most of it away.
Flux 2 Klein
$0.008 per image
Maker
Black Forest Labs
Takes
text-to-image
Custom LoRA
No
Most persona work has been built on the Flux family. Klein is the cheap tier of the current generation.
GPT Image 2.5 Edit
$0.039 per generation
Maker
OpenAI
Takes
image-edit
Custom LoRA
No
OpenAI's newest image model (September 2026), also in ChatGPT. Very realistic from reference images of your own persona. Price for 1K at medium quality with one reference image; each extra reference adds about $0.015.
Seedream 4.5
$0.04 per image
Maker
ByteDance
Takes
text-to-image
Custom LoRA
No
ByteDance’s image model. Its price sits between the cheap bulk models and the premium ones.
Seedream V5.0 Pro Edit
$0.045 per generation
Maker
ByteDance
Takes
image-edit
Custom LoRA
No
Edits one or more images you give it and does not generate from nothing. This is the mode you need to fix a single frame. On 5 October 2026 WaveSpeed took 10% off, so one edit cost $0.0405.
GPT Image 2 Edit
$0.07 per generation
Maker
OpenAI
Takes
image-edit
Custom LoRA
No
Edits an image from instructions in normal language. Good at text inside an image, which most models still get wrong. On 5 October 2026 WaveSpeed took 5% off, so one edit cost $0.0665.
Nano Banana 2
$0.07 per image
Maker
Google
Takes
text-to-image, image-edit
Custom LoRA
No
The standard tier of Google’s image models. Price at 1K. On 5 October 2026 WaveSpeed took 10% off, so one image cost $0.063.
Nano Banana Pro
$0.14 per image
Maker
Google
Takes
text-to-image, image-edit
Custom LoRA
No
The premium tier. Most subscription credit systems base their prices on this model. On 5 October 2026 WaveSpeed took 10% off, so one image cost $0.126.

What this table does not tell you

Whether the face stays the same. No vendor publishes a measure, a benchmark or even a definition of character consistency. It decides whether you can use a model for a persona, and no spec sheet shows it. Test each model on your own character.
Which row is cheaper. You cannot compare a per-second price with a per-generation price directly. For a ten-second clip, a model at $0.10 per second costs ten times as much as a model at $0.10 per clip.
What you will really spend. These are list prices for one output. Your real cost depends on how many outputs you throw away, and nobody publishes that number either.
Licensing. Open weights are not the same as a commercial licence. Several models with published weights limit commercial use. Read the terms of a model before you earn money with it.
Whether these are the best models. These are the models that matter for persona work on the two hosts this site refers to. Those hosts offer about a thousand models in total.

Sources

The same figures are available as machine-readable JSON. It is rebuilt together with the table, so you can work with the numbers without scraping this page.

Which models hold the face explains how to read all of this for a persona, and why the mode matters more than the model. What an AI persona costs to start shows what you spend before the account earns.

5,856+ members

Got a question about this? Ask it in #technical-inquiries on our Discord. People who run AI influencers answer there. It is free, and you can read everything right away.

Join the Discord What is inside the server