Image and Video Generation APIs Compared (2026)
/ Arvid Andersson
TLDR: Most image and video APIs now sell the same popular models. What differs is the billing unit, the resolution tiers and whose terms you sign. If you want many models behind one key, look at a multi-model platform like fal, Replicate, Together or Runware. If you've settled on one model family, the lab's own API is usually the straighter path.
This week I added a batch of image and video generation APIs to Infrabase and went through each vendor's pricing page. The thing is, the same model shows up almost everywhere and almost nowhere is it priced the same way.
Veo 3.1 Fast costs $0.10 to $0.12 a second from Google, depending on resolution. Through Runway it costs 15 credits, which is $0.15 a second. Seedance 2.5 is billed per second on two platforms and per video on a third, without saying how long the video is. Every price here is from the vendor's own page on October 11, 2026. I haven't converted between units, since per second, per video, per image and per token only line up if you guess at things the vendors don't publish.
Three ways to get the same model
There's also a fourth thing that gets called video generation: a live avatar you talk to. It's a different purchase and gets its own card.
A multi-model API
fal, Replicate, Together AI, Runware
Fits when You want to try several models, or switch later, without new accounts.
Check first The billing unit for the exact model you plan to use. It varies within one platform.
The lab's own API
Runway, Black Forest Labs, Luma, LTX, Recraft
Fits when You've picked a model family and want its full feature set.
Check first That one account per lab is fine for you. Read the lab's terms too.
Your own weights on serverless GPUs
Modal, Baseten, Cerebrium
Fits when You run an open model (maybe fine-tuned) and want to pay for compute time rather than per output.
Check first The licence on the weights. Open doesn't always mean commercial.
A live avatar
Tavus, Anam, LemonSlice
Fits when You want a talking video agent in a session, not a generated clip.
Check first The minutes included in each plan.
The same model on different platforms
Veo 3.1 (Google), per second with audio
| Standard | Fast | |
|---|---|---|
Google Gemini API
|
$0.40
720p and 1080p; $0.60 at 4K
|
$0.10
720p; $0.12 at 1080p
|
Runway
|
$0.40
40 credits
|
$0.15
15 credits
|
Runway sells credits at $0.01 each. Google also has Veo 3.1 Lite at $0.05 a second for 720p. fal lists Veo 3.1 but doesn't show a public price.
Seedance 2.5 (ByteDance)
| Price | Unit | |
|---|---|---|
Runware
|
$0.1025
480p
|
per second
|
|
Runway
|
$0.20
20 credits at 480p, 30 at 720p, 68 at 1080p
|
per second
plus input video, 80-credit minimum
|
Together AI
|
$0.115
|
per video
length and resolution not stated
|
The per-second prices compare directly. The per-video one doesn't until you know the default length. I'd ask before budgeting on it.
Multi-model platforms
These sell many labs' models through one key and one bill. Some also serve LLMs on the same key, which is handy if your product writes text and makes images in the same flow.
- fal lists more than 1,000 image, video, audio and 3D models. It also rents serverless GPUs and GPU compute for your own models. Prices are per model and the units vary: Kling v3 Pro is $0.14 a second, Seedance 2.0 is billed per 1,000 tokens.
- Replicate runs and fine-tunes open-source models through one API. You can deploy your own models on the same platform.
- Together AI has FLUX 3 for images and a long video list (Seedance, Kling, Sora 2, Veo, Hailuo), priced per video, next to its LLMs.
- Runware puts image, video, audio, language and 3D models on one endpoint with the same auth and billing. It has an MCP server too.
- SiliconFlow prices FLUX images per image and Wan videos per video, next to its chat models.
- DigitalOcean Inference Engine serves OpenAI, Anthropic and open-weight LLMs plus image models hosted by fal, all from one prepaid balance. FLUX Schnell is $0.003 per megapixel.
If you already use Gemini or OpenAI for text, both generate images through the same API. Google prices Nano Banana 2.1 at $30 per 1M output tokens and spells that out as $0.0336 for a 1K image or $0.113 at 4K. OpenAI bills gpt-image-2.5 per token too ($8 input, $30 output per 1M image tokens) but doesn't give a per-image figure.
The labs' own APIs
| Makes | From | Unit | |
|---|---|---|---|
|
Runway
|
Video: Gen-4 Turbo, Gen-4.5, Aleph 2
also resells Veo, Seedance and Wan
|
$0.05
Gen-4 Turbo
|
per second
|
Black Forest Labs
|
Open weights
FLUX images, FLUX 3 Video
weights under a non-commercial licence
|
$0.041
FLUX 3 Image, 768x768
|
per image
FLUX.2 per megapixel
|
Luma AI
|
Ray3.2 video, Uni-1.1 images
|
$0.15
5 s at 540p
|
per video
|
LTX
|
LTX-2.5 video up to 4K
|
$0.09
fast, 720p
|
per second
|
Recraft API
|
SVG
Images for design work, raster or vector
|
$0.007
V4.1 Flash
|
per image
|
Resolution changes these prices a lot. A 5-second Luma clip is $0.15 at 540p and $1.20 at 1080p. FLUX 3 Image goes from $0.041 at 768x768 to $0.607 at 4K. Krea API and ModelsLab are also listed on Infrabase. Krea's API prices are behind its app login and not in the table.
Live avatars
These sell a real-time video conversation, priced by the minute.
| Free | Paid from | At volume | |
|---|---|---|---|
Tavus
|
20 min/month
|
$22/month
60 minutes
|
$975/month
4,000 minutes
|
Anam
|
30 min/month
|
$0.16/min
over the plan
|
$0.04/min
"at scale"
|
LemonSlice
|
โ
|
$8/month
|
$0.048/min
"as low as"
|
What to check before you pick
- The unit, and what it covers. If one platform says per video and another says per second, find out the default length first.
- Resolution. The jump differs a lot. Veo 3.1 Standard costs the same at 720p and 1080p, LTX-2.5 Fast goes from $0.09 to $0.13 a second and a Luma clip from $0.30 to $1.20. HDR doubles Luma's price again.
- Minimums. Runway charges at least 80 credits per Seedance 2.5 generation and 56 for Aleph 2.
- Whether credits expire. Recraft's API units don't. Check before you buy a big block anywhere else.
- The licence, if you self-host. Black Forest Labs publishes several FLUX weights under a non-commercial licence and sells a separate licence for commercial self-hosting.
- MCP. Runway, Recraft, Krea and Runware have MCP servers. You can call them from Claude or another assistant.
Prices in this space move fast. Treat the numbers here as a snapshot of October 2026. None of these products pays for its listing on Infrabase. If a provider spots something wrong, I'll correct it here.
Resources
- Media Generation: every image and video API listed on Infrabase
- Inference API comparison: LLM providers compared on price
- Pricing pages used: Google Gemini API, Runway, Together AI, Runware, fal, Black Forest Labs, Luma, LTX, Recraft, DigitalOcean, Tavus, Anam, LemonSlice
Is your product missing?