AI & ML
Image and Video Generation APIs Compared (2026): fal, Replicate, Runware and More
Seven image and video generation APIs for developers: how each one bills, example prices, models, async jobs and webhooks, JS/TS SDKs and free credits.
To add image or video generation to an app, pick between a multi-model platform and a model maker's own API. fal, Replicate, Runware and Together AI host many models behind one key, each with a different unit of billing. Black Forest Labs, Google and OpenAI sell their own models directly. For video, note that two big names are gone: OpenAI's Sora API and Veo 3.1 Preview in the Gemini API have both been shut down.
Seven providers are below, checked against each vendor's own pricing page and docs on 11 November 2026. Prices are examples from those pages, not a full price list. Model catalogs and prices change often, so check the page before you design around a number.
What to compare
Four things matter more than the model list when you wire one of these into an app:
- The billing unit. Per image, per megapixel, per second of output video, per second of GPU time, or per token. The same picture can cost very different amounts depending on resolution and quality settings.
- Async jobs. Video and large images take seconds to minutes. A request that blocks until the file is ready will not fit in a serverless function's timeout.
- Webhooks. Without them you poll. With them, the provider calls your server when the job is done, so you need a public HTTPS endpoint and a handler that can take duplicates.
- Free credits. Few of these give any. The ones that do are listed; where a provider does not publish an amount, the entry says so.
Multi-model platforms
One API key, many models from different labs.
fal
fal prices each model by its output: per megapixel or per image for most image models, per second for video. Examples from the pricing page: FLUX.1 [dev] at $0.025 per megapixel, Nano Banana Pro at $0.15 per image, and Kling Video v2.6 Pro image-to-video at $0.07 per second. It also rents serverless GPUs billed per second, with an H100 listed at $4.50 an hour.
Queued requests are the
recommended way to call models. Submit with fal.queue.submit in the @fal-ai/client
package, then poll or pass webhookUrl (the fal_webhook query parameter over HTTP) to
get the result posted to your server. The FAQ says
new accounts start at 2 concurrent requests, rising to 40 as you buy credits. Free credits
and coupons exist, but no standard amount is published. The
FLUX.1 [dev] page lists commercial use through
fal's platform; check each model's page for its own terms.
Replicate
Replicate bills in three ways depending on the model: per second of hardware time, per output image, or per second of output video. Examples: black-forest-labs/flux-1.1-pro at $0.04 per image, flux-schnell at $3.00 per thousand images, and wan-2.1-i2v-480p at $0.09 per second of video. If you run your own model, an H100 is $0.001525 per second.
Predictions are async. The HTTP API takes a
webhook URL and a webhook_events_filter (start, output, logs, completed), or a
Prefer: wait header to hold the request open for up to 60 seconds. The JavaScript client
is the replicate npm package, where
replicate.run waits for output and predictions.create returns at once. Replicate's
Try for Free collection lets you run
selected models free, with an unpublished cap.
Runware
Runware charges per image or per second of video, and only for successful requests that produce output. Examples: Recraft V4.1 Flash at $0.007 per image, Bria FIBO 2.0 at $0.04 per image, and Kling VIDEO 4.0 at $0.157 per second at 720p. The pricing page says exact cost depends on your parameters and points you to the Playground to estimate a request.
The API is REST, and the
TypeScript SDK (@runware/sdk) wraps both REST and WebSocket transports. Long tasks such
as video run asynchronously. Add webhookURL to a request and Runware
posts the result when it finishes; one call
per item in a batch, and possibly more than once, so track task UUIDs. New users who sign
up with a business email get $2 in free credits.
Together AI (images and video)
Together AI is better known for LLMs, but it lists image and video models too. Images are priced per image or per megapixel: FLUX.2 [pro] at $0.03 per image, SD XL at $0.0019. Video is listed per video, for example Wan 2.7 Text-to-Video at $0.10.
Image calls are a direct
together.images.generate call in the together-ai package, returning a URL or base64.
Video is an async job: create it, then
poll videos.retrieve until it is completed or failed. The docs do not mention
webhooks. There is no free trial: the first credit
purchase is a minimum of $15.
Model makers' own APIs
Fewer models, sold by the lab that makes them.
Black Forest Labs (FLUX)
The BFL API sells FLUX models directly, priced in dollars with 1 credit = $0.01. FLUX 3 Image is $0.048 per image at 1K and FLUX 3 Image Klein $0.009. FLUX.2 [pro] is per megapixel: $0.03 for the first and $0.015 for each additional one. FLUX 3 Video starts at $0.15 per second in HD.
Every request is async: you get back an id and a polling_url and poll it until the
status is Ready. The integration guide
warns that generated images expire after 10 minutes, so copy the file to your own storage
straight away. The API schema has webhook_url and
webhook_secret on most FLUX.1 and FLUX.2 endpoints, but not on the FLUX 3 image and video
ones. The docs show raw HTTP calls, with no official JS SDK. Free credits are not published.
Google (Gemini API)
Google's image models in the Gemini API are the Nano Banana family, priced per image by resolution: Nano Banana 2.1 is $0.0336 per 1K image, Nano Banana Pro $0.134 per 1K or 2K image. Video is Gemini Omni Flash, estimated at about $0.07 per second at 720p. Imagen is listed as shut down, and the video docs say the Veo 3.1 Preview was shut down on 22 October 2026.
The Gemini API's free tier does not cover these: the pricing page marks image and video
models "Not available" on it. Image generation
returns the image in the response of interactions.create in @google/genai, and can
also run through the Batch API. Omni video returns
inline by default, from 3 to 15 seconds long, or as a hosted file you poll. Google states
that all generated images and videos carry an invisible SynthID watermark.
OpenAI (images only)
OpenAI's image models are priced per
token, and the image generation guide
converts that to per-image costs. GPT-Image-2 at 1024×1024 is $0.006 at low quality, $0.053
at medium and $0.211 at high; GPT-Image-1 Mini starts at $0.005. You call it through the
openai npm package, and can stream up to three partial images while it works.
There is no video. The video guide says the Sora 2 models and Videos API shut down on 24 September 2026. Webhooks and free credits for image generation are not mentioned on the pages we checked.
Side by side
| Provider | Billing unit | Example price | Async / webhooks | JS/TS SDK | Free credits |
|---|---|---|---|---|---|
| fal | Per MP, per image, per video second | FLUX.1 [dev] $0.025/MP | Queue + webhooks | @fal-ai/client | Not published |
| Replicate | Hardware second, per image, per video second | flux-1.1-pro $0.04/image | Async + webhooks | replicate | Try for Free models |
| Runware | Per image, per video second | Recraft V4.1 Flash $0.007/image | Async + webhookURL | @runware/sdk | $2 (business email) |
| Together AI | Per image, per MP, per video | FLUX.2 [pro] $0.03/image | Video jobs polled; no webhooks documented | together-ai | None; $15 minimum |
| Black Forest Labs | Per image, per MP, per video second | FLUX 3 Image 1K $0.048 | Polling; webhooks on most FLUX.1/.2 | None documented | Not published |
| Google Gemini API | Per image, per video second (est.) | Nano Banana 2.1 $0.0336/1K | Inline; Batch API; video file polling | @google/genai | Not on free tier |
| OpenAI | Per token (per-image table) | GPT-Image-2 low $0.006 | Streaming partial images | openai | Not mentioned |
How to choose
- You want to try many models without many accounts: fal, Replicate or Runware. Compare the unit: fal and Runware price the output, Replicate often prices the hardware second.
- You are on serverless functions: prefer a provider with webhooks (fal, Replicate, Runware, most BFL endpoints) so no function waits for a video to render.
- You need one specific model family: go to its maker. FLUX is on BFL's own API first; Nano Banana and Omni are on the Gemini API.
- You need video: fal, Replicate, Runware, Together, BFL (FLUX 3 Video) or Google (Omni Flash). OpenAI no longer offers it.
- You want to test before paying: Runware's $2, Replicate's free collection. Together and the Gemini API image and video models need paid credits.
- You ship to a commercial product: read the license on each model's page. fal states commercial use for FLUX.1 [dev] on its platform; other providers leave it to model terms.
FAQ
Which image generation API is cheapest?
It depends on the model and resolution, not just the provider. Among the prices quoted here, Together's SD XL ($0.0019 per image) and Replicate's flux-schnell ($3.00 per thousand) are the lowest; OpenAI's GPT-Image-2 at low quality is $0.006. Compare at the size and quality you will actually ship.
Is there a free image generation API?
Not much of one. Runware gives $2 to business-email signups and Replicate has a Try for Free collection with an unpublished cap. The Gemini API's free tier does not include image or video models. More one-off credits are on the free AI API offers page.
Can I still use Sora or Veo through an API?
Not from the original vendors. OpenAI's Sora Videos API shut down on 24 September 2026, and Veo 3.1 Preview in the Gemini API on 22 October 2026. Google points users to Gemini Omni Flash. Together AI still lists Veo 3.1 models on its pricing page.
Do I need webhooks?
For video, practically yes, unless you run a worker that can poll. fal, Replicate and Runware send webhooks; BFL does on most FLUX.1 and FLUX.2 endpoints. Replicate and Runware both warn that a webhook can arrive more than once, so make your handler idempotent.
How current are these prices?
Every figure was checked against the vendor's own page on 11 November 2026. Generation models change monthly, so treat the numbers as a starting point.
For text models with a free tier that lasts, see free AI API tiers. Current free and trial credits, including media APIs, are on the offers page. For more tools to build with, browse the AI & ML directory.